JP2004272292A - Sound signal processing method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method and an apparatus to process the inputted sound signals which include degraded sound such as quantizing noise so that the degraded sound is subjectively unperceptible. <P>SOLUTION: The spectrum of a decoded sound after performing auditory weighting to the decoded sound as the input sound signals, is computed and transformation strength is calculated on the basis of the magnitude of the amplitude and the continuity of the spectrum. In a signal transformation section, the spectrum of the decoded sound is obtained, amplitude smoothing and phase disturbance adding are conducted on the basis of the transformation strength and the spectrum is returned to a signal region to provide transformed and decoded voice. In a signal evaluation section, the decoded sound is analyzed to obtain background noise likeness and the obtained value is made an addition control value. In a weighted value adding section, the weight for adding to the decoded sound is reduced, the weight for adding to the transformed sound is increased, and an output sound is obtained, when the addition control value indicates the background noise likeness. <P>COPYRIGHT: (C)2004,JPO&NCIPI

Description

  The present invention makes it difficult to subjectively perceive subjectively unfavorable components such as quantization noise generated by encoding / decoding processes such as voices and musical sounds, distortion caused by various signal processing processes such as noise suppression processing, and the like. Signal processing method and a sound signal processing apparatus for performing the above processing.

  As the compression rate of information source coding such as voice and musical sound is increased, quantization noise, which is distortion at the time of coding, gradually increases, and the quantization noise is deformed and becomes subjectively unbearable. Come. For example, in the case of a speech coding system that attempts to faithfully represent a signal itself, such as PCM (Pulse Code Modulation) and ADPCM (Adaptive Differential Pulse Code Modulation), the quantization noise is random. Yes, it does not matter much subjectively, but as the compression ratio increases and the coding system becomes more complex, the quantization noise may show spectral characteristics unique to the coding system, resulting in a subjectively large deterioration. Come. In particular, in a signal section in which background noise is dominant, a speech model used by a speech encoding scheme with a high compression rate does not match, resulting in a very hard-to-hear sound.

  In addition, when noise suppression processing such as a spectral subtraction method is performed, noise estimation errors remain as distortions on the processed signal, and have significantly different characteristics from the signal before the processing, which greatly deteriorates the subjective evaluation. Sometimes.

  As a conventional method for suppressing the deterioration of the subjective evaluation due to the quantization noise and distortion as described above, JP-A-8-130513, JP-A-8-146998, JP-A-7-160296, and JP-A-6-326670 are disclosed. And JP-A-7-248793, and S.I. F. Boll, fractionSSP-27, No. 2, pp. 113-120, April 1979) (hereinafter referred to as Document 1).

  Japanese Patent Application Laid-Open No. Hei 8-130513 aims to improve the quality of a background noise section, and determines whether or not the section is only a background noise section, and performs an encoding process or decoding dedicated to the section including only the background noise section. By performing the processing and suppressing the characteristic of the synthesis filter when decoding only the section of the background noise, an acoustically natural reproduced sound is obtained.

  Japanese Patent Application Laid-Open No. Hei 8-146998 discloses a technique of adding white noise or pre-stored background noise to decoded speech with the aim of suppressing white noise from becoming an unpleasant tone due to encoding and decoding. It was done.

  Japanese Patent Application Laid-Open No. 7-160296 seeks an auditory masking threshold based on an index related to a decoded speech or a spectrum parameter received by a speech decoding unit and aims to reduce quantization noise audibly, and reflects this. A filter coefficient is obtained, and this coefficient is used for a post filter.

  Japanese Unexamined Patent Publication No. Hei 6-326670 discloses a system in which code transmission is stopped in a section that does not include voice for communication power control or the like. In the absence of code transmission, pseudo background noise is generated and output on the decoding side. It aims to reduce the sense of discomfort between the actual background noise included in the voice section and the pseudo background noise in the silent section, and superimposes the pseudo background noise not only on the section containing no voice but also on the voice section. It is intended to be.

  Japanese Patent Application Laid-Open No. Hei 7-248793 aims to reduce audibly the distortion sound generated by the noise suppression processing. On the encoding side, first, it is determined whether the noise section is a speech section or a speech section. Is transmitted in the voice section, the spectrum after the noise suppression processing is transmitted, and the decoding side generates and outputs a synthesized sound using the received noise spectrum in the noise section. Is added to the synthesized sound generated by using the spectrum of (i) and the synthesized sound generated by using the noise spectrum received in the noise section, multiplied by the superimposition ratio, and output.

Literature 1 aims to reduce the distorted sound generated by the noise suppression processing audibly, and performs smoothing on the temporally preceding and succeeding sections and the amplitude spectrum of the output voice after the noise suppression processing. The amplitude suppression processing is performed only in the background noise section.
JP-A-8-130513 JP-A-8-146998 JP-A-7-160296 JP-A-6-326670 JP-A-7-248793 S. F. Boll, fractionSSP-27, No. 2, pp. 113-120, April 1979).

  The above conventional method has the following problems.

  Japanese Patent Application Laid-Open No. Hei 8-130513 has a problem that a sudden change in characteristics occurs at a boundary between a noise section and a speech section because encoding processing and decoding processing are largely switched according to the section determination result. In particular, when erroneous determination of a noise section as a voice section frequently occurs, the noise section which is originally relatively stationary fluctuates in an unstable manner, and may rather deteriorate the noise section. When transmitting the noise section determination result, it is necessary to add information to be transmitted, and there is a problem that when the information is erroneous on a transmission path, unnecessary deterioration is caused. Further, simply suppressing the characteristics of the synthesis filter does not reduce the quantization noise generated at the time of excitation coding, so that there is a problem that an improvement effect is hardly obtained depending on the type of noise.

  Japanese Patent Application Laid-Open No. 8-146998 has a problem that the characteristic of the current coded background noise is lost because noise prepared in advance is added. In order to make the degraded sound difficult to hear, it is necessary to add noise at a level higher than the degraded sound, and there is a problem that the reproduced background noise increases.

  In Japanese Patent Application Laid-Open No. 7-160296, an auditory masking threshold is obtained based on spectral parameters, and a spectrum post-filter is simply performed based on the threshold. There is a problem that there is hardly any improvement effect. In addition, since a large change cannot be given to a main component that is not masked, there is a problem that no improvement effect can be obtained for distortion included in the main component.

  In JP-A-6-326670, since the pseudo background noise is generated regardless of the actual background noise, there is a problem that the characteristics of the actual background noise are lost.

  Japanese Patent Laid-Open No. Hei 7-248793 has a problem that since the encoding process and the decoding process are largely switched in accordance with the section determination result, erroneous determination of a noise section or a voice section causes a large deterioration. If a part of the noise section is mistaken for a voice section, the sound quality in the noise section varies discontinuously, making it difficult to hear. Conversely, if the speech section is mistaken for a noise section, speech components are mixed into the synthesized sound of the noise section using the average noise spectrum and the synthesized sound using the noise spectrum superimposed in the speech section, and There is a problem that sound quality degradation occurs. Furthermore, in order to make the degraded sound in the voice section inaudible, it is necessary to superimpose noise that is not low.

  Literature 1 has a problem that a processing delay of a half section (about 10 ms to 20 ms) occurs due to smoothing. In addition, when a part of the noise section is erroneously determined to be a speech section, there is a problem that the sound quality in the noise section varies discontinuously, making it difficult to hear.

  The present invention has been made in order to solve such a problem, and there is little deterioration due to a section determination error, little dependence on a noise type or a spectrum shape, and a large delay time is not required. Sound signal processing that can retain the characteristics, does not excessively increase the background noise level, does not require the addition of new transmission information, and can provide a good suppression effect even for components degraded by excitation coding etc. It is an object to provide a method and a sound signal processing device.

  The input sound signal is processed to generate a first processed signal, the input sound signal is analyzed to calculate a predetermined evaluation value, and the input sound signal and the first processed signal are calculated based on the evaluation value. The second processing signal is obtained by weighting and adding, and the second processing signal is used as an output signal.

  Further, the first processed signal generation method calculates a spectral component for each frequency by performing a Fourier transform on the input sound signal, and a predetermined spectral component for each frequency calculated by the Fourier transform. It is characterized in that a transform is given and the transformed spectral component is generated by performing an inverse Fourier transform.

  Further, the weighted addition is performed in a spectral domain.

  Further, the weighted addition is controlled independently for each frequency component.

  Further, the predetermined deformation of the spectrum component for each frequency includes a smoothing process of an amplitude spectrum component.

  Furthermore, the predetermined deformation of the spectrum component for each frequency includes a disturbance imparting process of a phase spectrum component.

  Further, the smoothing strength in the smoothing processing is controlled by the magnitude of the amplitude spectrum component of the input sound signal.

  Further, the present invention is characterized in that the disturbance imparting strength in the disturbance imparting process is controlled by the magnitude of the amplitude spectrum component of the input sound signal.

  Further, the smoothing strength in the smoothing process is controlled by the magnitude of the continuity of the spectral components of the input sound signal in the time direction.

  Further, the disturbance imparting strength in the disturbance imparting process is controlled by the magnitude of the continuity of the spectral components of the input sound signal in the time direction.

  Further, an input sound signal weighted by auditory sense is used as the input sound signal.

  Further, the smoothing strength in the smoothing process is controlled by the magnitude of the time variability of the evaluation value.

  Further, the disturbance imparting intensity in the disturbance imparting process is controlled by the magnitude of the time variability of the evaluation value.

  Further, a characteristic of the background noise calculated by analyzing the input sound signal is used as the predetermined evaluation value.

  Further, the method is characterized in that a degree of fricativeness calculated by analyzing the input sound signal is used as the predetermined evaluation value.

  Further, a decoded speech obtained by decoding a speech code generated by a speech encoding process is used as the input sound signal.

  In the sound signal processing method of the present invention, the input sound signal is used as a first decoded sound obtained by decoding a sound code generated by a sound coding process, and the first decoded sound is subjected to a post-filter process to perform a second filtering. The first decoded voice is generated to generate a first processed voice, a decoded voice is analyzed to calculate a predetermined evaluation value, and based on the evaluation value, A second decoded audio and the first processed audio are weighted and added to form a second processed audio, and the second processed audio is output as an output audio.

  A sound signal processing device according to the present invention includes a first processed signal generation unit that processes an input sound signal to generate a first processed signal, and an evaluation value that analyzes the input sound signal and calculates a predetermined evaluation value. A second processing signal generator that weights and adds the input sound signal and the first processing signal based on the evaluation value of the evaluation value calculator and outputs the second processing signal as a second processing signal. It is characterized by having.

  Further, the first processed signal generation unit calculates a spectrum component for each frequency by performing a Fourier transform on the input sound signal, and smoothes an amplitude spectrum component with respect to the calculated spectrum component for each frequency. And a first processing signal is generated by performing an inverse Fourier transform on the spectrum component after the smoothing process of the amplitude spectrum component.

  Further, the first processed signal generation unit calculates a spectral component for each frequency by performing a Fourier transform on the input sound signal, and disturbs a phase spectral component with respect to the calculated spectral component for each frequency. An application process is performed, and the spectrum component after the disturbance application process of the phase spectrum component is subjected to inverse Fourier transform to generate a first processed signal.

  As described above, the sound signal processing method and the sound signal processing apparatus of the present invention perform a predetermined signal processing process on an input signal so that a deterioration component included in the input signal is not subjectively noticed. Generated processing signal, and the addition weight of the input signal and the processing signal is controlled by a predetermined evaluation value, so that the ratio of the processing signal is increased mainly in the section where many degraded components are included, and the subjective quality is improved. There is an effect that can be done.

  In addition, the conventional binary interval determination is eliminated, and the continuous value evaluation value is calculated. Based on this, the weighted addition coefficient of the input signal and the processed signal can be controlled continuously, so that quality deterioration due to an interval determination error is avoided. There is an effect that can be done.

  In addition, since an output signal can be generated by processing an input signal that contains a lot of background noise information, a stable quality improvement effect that does not depend much on the noise type or spectrum shape while maintaining the characteristics of the actual background noise can be obtained. It is possible to obtain an effect of improving a component degraded by excitation coding or the like.

  In addition, since processing can be performed using the input signal up to the present time, a particularly large delay time is not required, and there is an effect that a delay other than the processing time can be eliminated depending on a method of adding the input signal and the processed signal. If the level of the input signal is lowered when raising the level of the processing signal, it is not necessary to superimpose large pseudo noise to mask the degraded components as in the past, and conversely, depending on the application target Thus, it is possible to make the background noise level small or even large. Also, needless to say, even in the case of eliminating the degraded sound due to the voice encoding / decoding, it is not necessary to add new transmission information as in the related art.

  The sound signal processing method and the sound signal processing apparatus of the present invention perform predetermined processing in the spectral domain on the input signal, so that the deterioration component included in the input signal is not subjectively noticed. Since the processing signal is generated and the addition weight of the input signal and the processing signal is controlled by a predetermined evaluation value, in addition to the effect of the signal processing method, a process of suppressing a fine degradation component in a spectrum region is performed. Has the effect of further improving the subjective quality.

  According to the sound signal processing method of the present invention, in the sound signal processing method of the present invention, the input signal and the processed signal are weighted and added in the spectral domain. For example, in the case of connecting to the subsequent stage of the noise suppression method that performs the processing of the above, some or all of the Fourier transform processing and the inverse Fourier transform processing required by the sound signal processing method can be omitted, and the effect of simplifying the processing can be obtained. is there.

  According to the sound signal processing method of the present invention, in the sound signal processing method of the present invention, the weighted addition is controlled independently for each frequency component. The dominant component of the degraded component is replaced with the processed signal with emphasis, and it is no longer replaced with a good component with a small amount of quantization noise and degraded components. There is an effect that the deteriorating component can be suppressed subjectively and the subjective quality can be improved.

  The sound signal processing method of the present invention performs the smoothing processing of the amplitude spectrum component as the processing in the sound signal processing method of the present invention. As a result, it is possible to favorably suppress unstable fluctuations of the amplitude spectrum component caused by the above-mentioned factors, and to improve the subjective quality.

  According to the sound signal processing method of the present invention, as the processing in the sound signal processing method of the present invention, the disturbance imparting process of the phase spectrum component is performed. The quantization noise and the degradation components, which often have characteristic correlation with each other, can disturb the relationship between the phase components and improve the subjective quality. is there.

  In the sound signal processing method of the present invention, the smoothing strength or the disturbance imparting strength in the sound signal processing method of the present invention is controlled by the magnitude of the amplitude spectrum component of the input signal or the input signal that is auditory weighted. In addition to the effects of the sound signal processing method, processing is added with emphasis on components in which the quantization noise and degraded components are dominant because the amplitude spectrum components are small, and quantization noise and degraded components are added. This eliminates the possibility of processing even a good component with a small amount of noise, has the effect of subjectively suppressing quantization noise and degraded components while maintaining the characteristics of the input signal, and improving the subjective quality.

  According to the sound signal processing method of the present invention, the smoothing strength or the disturbance imparting strength in the sound signal processing method of the present invention is controlled by the magnitude of the temporal continuity of the spectral component of the input signal or the auditory weighted input signal. Therefore, in addition to the effects of the above sound signal processing method, processing is added with emphasis on components that tend to increase quantization noise and degraded components due to the low continuity of spectral components. There is no need to process even a good component with little noise and degraded components, and it is possible to subjectively suppress quantization noise and degraded components while maintaining good characteristics of the input signal, thereby improving the subjective quality.

  According to the sound signal processing method of the present invention, the smoothing strength or the disturbance imparting strength in the sound signal processing method of the present invention is controlled by the magnitude of the time variability of the evaluation value. In addition to the effects, it is possible to suppress unnecessarily strong processing in a section where the characteristics of the input signal are fluctuating, and it is possible to prevent the occurrence of echo and the occurrence of echoes, particularly by amplitude smoothing.

  The sound signal processing method of the present invention uses the degree of background noise likeness as the predetermined evaluation value in the sound signal processing method of the present invention. Prioritized processing is applied to background noise sections where deterioration components tend to occur frequently, and appropriate processing (no processing, low-level processing, etc.) is selected for sections other than background noise. Therefore, there is an effect that the subjective quality can be improved.

  The sound signal processing method of the present invention uses the degree of fricative likeness as the predetermined evaluation value in the sound signal processing method of the present invention.In addition to the effects of the sound signal processing method, quantization noise and noise Focused processing is applied to the friction noise section where a lot of deterioration components tend to occur, and appropriate processing (no processing, low-level processing, etc.) is selected for the section other than the friction sound as well. This has the effect of improving the subjective quality.

  The sound signal processing method of the present invention uses a sound code generated by a sound coding process as an input, decodes the sound code to generate a decoded sound, and uses the decoded sound as an input to use the sound signal processing method. Since the processed voice is generated by performing the signal processing and the processed voice is output as the output voice, the effect that the voice decoding having the subjective quality improvement effect and the like of the above-described sound signal processing method can be realized is realized. is there.

  The sound signal processing method of the present invention receives a speech code generated by a speech encoding process as an input, decodes the speech code to generate a decoded speech, performs a predetermined signal processing process on the decoded speech, and outputs the processed speech. Generate and perform post-filter processing on the decoded voice, further analyze the decoded voice before or after the post-filter, calculate a predetermined evaluation value, and weight the decoded voice after the post-filter and the processed voice based on this evaluation value. Since the addition and the output are performed, in addition to the effect of realizing the audio decoding having the subjective quality improvement effect and the like of the sound signal processing method as described above, it is possible to generate the processed audio which is not affected by the post filter. Since it is possible to perform highly accurate addition weight control based on the highly accurate evaluation value calculated without being affected by the above, there is an effect that the subjective quality is further improved.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

Embodiment 1 FIG.
FIG. 1 shows an overall configuration of a speech decoding method to which a sound signal processing method according to the present embodiment is applied. In FIG. 1, 1 is a speech decoding device, 2 is a signal processing unit that executes the signal processing method according to the present invention, and 3 is a signal processing unit. The voice code, 4 is a voice decoding unit, 5 is a decoded voice, and 6 is an output voice. The signal processing unit 2 includes a signal transformation unit 7, a signal evaluation unit 12, and a weighted addition unit 18. The signal transformation unit 7 includes a Fourier transform unit 8, an amplitude smoothing unit 9, a phase disturbance unit 10, and an inverse Fourier unit 11. The signal evaluation unit 12 includes an inverse filter unit 13, a power calculation unit 14, a background noise likeness calculation unit 15, an estimated background noise power update unit 16, and an estimated noise spectrum update unit 17.

  The operation will be described below with reference to the drawings.

  First, the speech code 3 is input to the speech decoding unit 4 in the speech decoding device 1. The voice code 3 is output as a result of separately coding a voice signal by a voice coding unit, and is input to the voice decoding unit 4 via a communication path or a storage device.

  The audio decoding unit 4 performs a decoding process on the audio code 3 in a pair with the audio encoding unit, and outputs a signal of a predetermined length (one frame length) obtained as a decoded audio 5. Then, the decoded speech 5 is input to the signal transformation unit 7, the signal evaluation unit 12, and the weighted addition unit 18 in the signal processing unit 2.

  The Fourier transform unit 8 in the signal transformation unit 7 performs windowing on the signal obtained by combining the input decoded speech 5 of the current frame and the latest portion of the decoded speech 5 of the previous frame as necessary, and performs post-windowing. By performing a Fourier transform process on the signal of (i), a spectral component for each frequency is calculated, and this is output to the amplitude smoothing unit 9. Note that typical examples of the Fourier transform processing include a discrete Fourier transform (DFT) and a fast Fourier transform (FFT). As the windowing process, various types such as trapezoidal windows, rectangular windows, Hanning windows can be applied, but here, a modified trapezoidal window in which the inclined parts at both ends of the trapezoidal window are replaced with half of the Hanning window respectively I do. An example of the actual shape and the time relationship with the decoded speech 5 and the output speech 6 will be described later with reference to the drawings.

  The amplitude smoothing unit 9 performs a smoothing process on the amplitude component of the spectrum for each frequency input from the Fourier transform unit 8 and outputs the smoothed spectrum to the phase disturbance unit 10. Regarding the smoothing process used here, the effect of suppressing degraded sound such as quantization noise can be obtained regardless of whether it is used in the frequency axis direction or the time axis direction. However, if the smoothing in the frequency axis direction is too strong, the spectrum will be sluggish and the characteristic of the original background noise will often be lost. On the other hand, if the smoothing in the time axis direction is made too strong, the same sound will remain for a long time, and a feeling of reverberation will be generated. As a result of making adjustments for various background noises, the quality of the output speech 6 was good when smoothing was not performed in the frequency axis direction and amplitude was smoothed in the logarithmic domain in the time axis direction. The smoothing method at that time is represented by the following equation.

y _i = y _i−1 (1−α) + x _i α Equation 1
Here, the logarithmic amplitude spectrum value before smoothing x _i is the current frame (i-th frame), y _i-1 is the previous frame logarithmic amplitude spectrum value after smoothing (the i-1 frame), y _i Is the logarithmic amplitude spectrum value of the current frame (i-th frame) after smoothing, α is a smoothing coefficient having a value of 0 to 1, and the smoothing coefficient α depends on the frame length, the level of the degraded sound to be eliminated, and the like. Although the optimum value is different, the value is generally about 0.5.

  The phase disturbance unit 10 disturbs the phase component of the smoothed spectrum input from the amplitude smoothing unit 9 and outputs the disturbed spectrum to the inverse Fourier transform unit 11. As a method of giving a disturbance to each phase component, a phase angle in a predetermined range is generated using random numbers, and this may be added to the original phase angle. If there is no restriction on the range of the phase angle generation, it is sufficient to simply replace each phase component with a phase angle generated by a random number. If the deterioration due to encoding or the like is large, the range of phase angle generation is not limited.

  The inverse Fourier transform unit 11 performs an inverse Fourier transform process on the disturbed spectrum input from the phase disturbing unit 10 to return the signal to the signal area and open a window for smooth connection with the preceding and succeeding frames. While performing the connection, the obtained signal is output to the weighted addition unit 18 as the modified decoded speech 34.

  The inverse filter unit 13 in the signal evaluation unit 12 performs an inverse filter process on the decoded speech 5 input from the speech decoding unit 4 using an estimated noise spectrum parameter stored in an estimated noise spectrum update unit 17 described later. And outputs the inverse-filtered decoded speech to the power calculator 14. By this inverse filter processing, the amplitude of the background noise is large, that is, the amplitude of a component that is highly likely to be antagonized between the voice and the background noise is suppressed. The signal power ratio in the background noise section can be made large.

  Note that the estimated noise spectrum parameter is selected from the viewpoints of compatibility with speech encoding processing and speech decoding processing and sharing of software. At present, a line spectrum pair (LSP) is often used. Similar effects can be obtained by using a spectral envelope parameter such as a linear prediction coefficient (LPC), a cepstrum, or the amplitude spectrum itself, in addition to the LSP. It is easy to use a linear interpolation or averaging process as the updating process in the estimated noise spectrum updating unit 17 described later, and it is guaranteed that the filter is stable even if the linear interpolation or the averaging process is performed among the spectral envelope parameters. Suitable LSPs and cepstrum are suitable. The cepstrum is superior in expressing the noise component to the spectrum, but the LSP is superior in terms of the easiness of the configuration of the inverse filter unit. When an amplitude spectrum is used, an LPC having this amplitude spectrum characteristic is calculated and used for an inverse filter, or the result of Fourier transform of the decoded speech 5 (equal to the output of the Fourier transform unit 8) is subjected to an amplitude transformation process. To achieve the same effect as the inverse filter.

  The power calculator 14 obtains the power of the inverse-filtered decoded speech input from the inverse filter 13 and outputs the calculated power value to the background noise likelihood calculator 15.

  The background noise likelihood calculating unit 15 uses the power input from the power calculating unit 14 and the estimated noise power stored in the estimated noise power updating unit 16 described later to calculate the background noise likelihood of the current decoded speech 5. The calculated value is output to the weighted addition unit 18 as the addition control value 35. Further, the calculated background noise likelihood is output to an estimated noise power updating unit 16 and an estimated noise spectrum updating unit 17 described later, and the power input from the power calculating unit 14 is output to the estimated noise power updating unit 16 described later. Output. Here, the likelihood of the background noise can be calculated most simply by the following equation.

_{v = log (p N) -} log (p) ··· formula 2
Here, p is the power input from the power calculator 14, p _N is the estimated noise power stored in the estimated noise power updater 16, and v is the calculated background noise likelihood.

In this case, the larger the value of v (the smaller the absolute value of a negative value, the smaller the absolute value), the more likely the background noise. In addition, various calculation methods such as calculating p _N / p to obtain v are conceivable.

  The estimated noise power updating unit 16 updates the estimated noise power stored therein using the background noise likelihood and the power input from the background noise likeness calculating unit 15. For example, when the likelihood of the input background noise is high (the value of v is large), the update is performed by reflecting the input power on the estimated noise power according to the following equation.

_{log (p N ') = (} 1-β) log (p N) + βlog (p) ··· Formula 3
Here, β is an update rate constant taking a value of 0 to 1 and may be set to a value relatively close to 0. The value on the right side of this equation is obtained, and updating is performed by setting p _N ′ on the left side as a new estimated noise power.

  In addition, regarding the method of updating the estimated noise power, the variability between frames is referred to in order to further improve the estimation accuracy, a plurality of input past powers are stored, and the noise power is determined by statistical analysis. Various modifications and improvements are possible, for example, by estimating the minimum value of p and using the lowest value of p as the estimated noise power.

  The estimated noise spectrum updating unit 17 first analyzes the input decoded speech 5 and calculates a spectrum parameter of the current frame. Spectral parameters to be calculated are as described in the inverse filter unit 13, and LSP is used in many cases. Then, the estimated noise spectrum stored therein is updated using the background noise likelihood input from the background noise likeness calculating unit 15 and the spectrum parameter calculated here. For example, when the likelihood of the input background noise is high (the value of v is large), the update is performed by reflecting the calculated spectral parameters in the estimated noise spectrum according to the following equation.

_{x N '= (1-γ} ) x N + γx ··· formula 4
Here, x is from the spectrum parameter of the current frame, x _N is the estimated noise spectrum (parameter). γ is an update rate constant taking a value of 0 to 1 and may be set to a value relatively close to 0. The value on the right side of this equation is obtained, and x _N ′ on the left side is updated as a new estimated noise spectrum (parameter).

  In addition, various improvements can be made to the method for updating the estimated noise spectrum in the same manner as the above-described method for updating the estimated noise power.

  Then, as the last processing, the weighted addition unit 18 receives the decoded speech 5 input from the speech decoding unit 4 and the input from the signal transformation unit 7 based on the addition control value 35 input from the signal evaluation unit 12. The modified decoded speech 34 is weighted and added, and the obtained output speech 6 is output. As the operation of the control method of weighted addition, as the addition control value 35 becomes larger (the likelihood of the background noise becomes higher), the weight for the decoded speech 5 is made smaller and the weight for the modified decoded speech 34 is made larger. Conversely, as the addition control value 35 becomes smaller (lower the likelihood of background noise), the weight for the decoded speech 5 is controlled to be larger, and the weight for the modified decoded speech 34 is controlled to be smaller.

  Note that, in order to suppress the quality deterioration of the output sound 6 due to a sudden change in weight between frames, it is desirable to perform smoothing so that the addition control value 35 or the weighting coefficient gradually changes for each sample.

  FIG. 2 shows a control example of weighted addition based on the addition control value in the weighted addition section 18.

In FIG. 2 (a), a case where linearly controlled using two threshold v ₁ and v ₂ for the addition control value 35. Addition control value 35 is the case of less than _{v 1,} the 1 weighting coefficient _{w S} for the decoded speech 5, the weighting coefficient _{w N} to deformation decoded speech 34 to 0. If the addition control value 35 is _{v 2} or more, the weighting coefficient _{w S} for the decoded speech 5 0, the weighting coefficient _{w N} to deformation decoded speech 34 to _{A N.} And when the addition control value 35 is _v less than _{2 v 1} or more, linear weighting coefficients _{w S} for the decoded speech 5 1-0, the weighting coefficient _{w N} to deformation decoded speech 34 between 0 to A _N Is calculated and given.

If this control is that as a reliably only modified decoded signal 34 when (v ₂ or more) that can be determined that the background noise interval is output, can be reliably determined that the speech segment (v less than ₁₎ Outputs the decoded speech 5 itself, and if it is not possible to determine whether it is a speech section or a background noise section (v ₁ or more and less than v ₂ ), the decoded speech 5 and the modified decoding at a ratio depending on which tendency is stronger. The result obtained by mixing the audio 34 is output.

Note that if you give a value of 1 or less as the weighting coefficient value A _N for multiplying the modified decoded signal 34 when (v ₂ or more) which can be determined that where a strictly background noise period, resulting in the amplitude suppression of the background noise period The effect is obtained. Conversely, if a value of 1 or more is given, an effect of enhancing the amplitude of the background noise section can be obtained. The amplitude of the background noise section often decreases due to the speech encoding / decoding process. In such a case, the reproducibility of the background noise can be improved by emphasizing the amplitude of the background noise section. Whether to perform amplitude suppression or amplitude emphasis depends on the application target, user requirements, and the like.

In FIG. 2 (b), is to add the new threshold value _{v 3, v} between ₁ and _{v 3, v 3} and _v if the weighting factor given by linearly calculated between _2. By adjusting the value of the weighting coefficient at the position of the threshold value v _3, it can be set more finely to the mixing ratio in the case of determining whether the speech segment or the background noise period is not attached (v ₁ or v less than _2). Generally, when two signals having low phase correlations are added, the power of the obtained signal is smaller than the sum of the powers of the two signals before the addition. v ₁ or v, 1 to the sum of the two weighting factors in the range of less than ₂ to be larger than w _N, it is possible to suppress the power reduction. The same effect can be obtained by taking a square root of the weighting coefficient obtained in FIG. 2A and further multiplying by a constant to obtain a new weighting coefficient.

In FIG. 2 (c), a value of B _N greater than 0 is given as a weighting coefficient w _{N to be applied} to the modified decoded speech 34 in a range less than v _{1 in} FIG. 2 (a), and accordingly, v ₁ or more and less than v ₂ This is a case where w _N in the range is also corrected. If the quantization noise or degraded sound in the voice section is large, such as when the background noise level is high or the compression rate in encoding is very high, even in the range where the voice section is surely known, By adding the modified decoded voice, it is possible to make the degraded sound difficult to hear.

FIG. 2D shows a control corresponding to a case where the background noise likelihood calculating unit 15 outputs the result (p _N / p) obtained by dividing the estimated noise power by the current power as the background noise likeness (addition control value 35). It is an example. In this case, since the addition control value 35 indicates the ratio of the background noise included in the decoded speech 5, the weighting coefficient is calculated so that the addition is performed at a ratio proportional to this value. Specifically, if _{w S} at _{w N} is 1 when the addition control value 35 is 1 or more is less than 0 and 1, _{w N} is the addition control value 35 itself, _{w S} is _{(1-w N} ).

    FIG. 3 is an explanatory diagram illustrating an example of an actual shape of a cutout window in the Fourier transform unit 8, a window for connection in the inverse Fourier transform unit 11, and a time relationship with the decoded speech 5.

    The decoded speech 5 is output from the speech decoding unit 4 at every predetermined time length (one frame length). Here, this one frame length is assumed to be N samples. FIG. 3A shows an example of the decoded speech 5, and x (0) to x (N-1) correspond to the decoded speech 5 of the input current frame. The Fourier transform unit 8 cuts out a signal of length (N + NX) by multiplying the decoded speech 5 shown in FIG. 3A by the modified trapezoidal window shown in FIG. 3B. NX is the length of each of the sections having a value less than 1 at both ends of the modified trapezoidal window. The sections at both ends are equal to a Hanning window having a length (2NX) divided into a first half and a second half. The inverse Fourier transform unit 11 multiplies the signal generated by the inverse Fourier transform process by the modified trapezoidal window shown in FIG. 3C, and obtains the signals in the preceding and succeeding frames (as indicated by broken lines in FIG. 3C). The signal is added while keeping the time relationship with the same signal, and a continuous modified decoded voice 34 (FIG. 3D) is generated.

  In the section (length NX) for connection with the signal of the next frame, the modified decoded speech 34 has not been determined at the time of the current frame. That is, the newly determined modified decoded speech 34 is x '(-NX) to x' (N-NX-1). Therefore, the output speech 6 obtained for the decoded speech 5 of the current frame is as follows.

y (n) = x (n) + x ′ (n) Equation 5
(N = -NX, ..., N-NX-1)
Here, y (n) is the output sound 6. At this time, the processing delay as the signal processing unit 2 requires at least NX.

  If the processing delay NX is an application target that cannot be tolerated, the output speech 6 can be generated as in the following equation by allowing the time lag between the decoded speech 5 and the modified decoded speech 34.

y (n) = x (n) + x '(n-NX) Expression 6
(N = 0, ..., N-1)
In this case, there is a time difference between the decoded speech 5 and the modified decoded speech 34, so that the disturbance in the phase disturbance unit 10 is weak (that is, the phase characteristics of the decoded speech remain to some extent), or the spectrum or power in the frame is not sufficient. When there is a sudden change, deterioration may occur. In particular, when the weighting coefficients in the weighted addition unit 18 change greatly, and when the two weighting coefficients are in opposition, deterioration is likely to occur. However, their deterioration is relatively small, and the effect of introducing the signal processing unit is sufficiently large. Therefore, this method can also be used for an application target for which the processing delay NX cannot be tolerated.

  In the case of FIG. 3, the modified trapezoidal window is multiplied before the Fourier transform and after the inverse Fourier transform, which may cause a decrease in the amplitude of the connected portion. This decrease in amplitude is likely to occur when the disturbance in the phase disturbance unit 10 is weak. In such a case, by changing the window before the Fourier transform to a square window, it is possible to suppress a decrease in amplitude. Normally, as a result of the phase being greatly deformed by the phase disturbance unit 10, the shape of the first deformed trapezoidal window does not appear in the signal after the inverse Fourier transform, so that a smooth connection with the deformed decoded speech 34 of the preceding and succeeding frames is obtained. Needs a second window.

  Here, the processing of the signal transformation unit 7, the signal evaluation unit 12, and the weighting addition unit 18 are all performed for each frame, but the present invention is not limited to this. For example, one frame is divided into a plurality of subframes, the processing of the signal evaluation unit 12 is performed for each subframe, an addition control value 35 for each subframe is calculated, and the weighting control in the weighting addition unit 18 is also performed for each subframe. You may go. Since the Fourier transform is used for the signal deformation processing, if the frame length is too short, the analysis result of the spectral characteristics becomes unstable, and the deformed decoded speech 34 is hardly stable. On the other hand, since the likelihood of the background noise can be calculated relatively stably even for a shorter section, a quality improvement effect in a rising part of voice or the like can be obtained by calculating for each subframe and finely controlling the weight.

  Further, the processing of the signal evaluation unit 12 may be performed for each sub-frame, and a small number of addition control values 35 may be calculated by combining all the addition control values in the frame. If it is not desired to make an error in the voice section as background noise, the minimum value (minimum value of the background noise likeness) of all the addition control values may be selected and output as the addition control value 35 representing the frame.

  Further, the frame length of the decoded voice 5 and the processing frame length of the signal transformation unit 7 do not need to be the same. For example, if the frame length of the decoded speech 5 is too short and too short for the spectrum analysis in the signal transformation unit 7, the decoded speech 5 of a plurality of frames may be accumulated and the signal transformation processing may be performed collectively. good. However, in this case, a processing delay occurs because the decoded speech 5 of a plurality of frames is accumulated. In addition, the processing frame length of the entire signal deforming unit 7 and the signal processing unit 2 may be set completely independently of the frame length of the decoded speech 5. In this case, the buffering of the signal becomes complicated, but the optimum processing frame length for the signal processing can be selected without depending on the frame length of various decoded voices 5, and the quality of the signal processing unit 2 becomes the best. effective.

    Also, here, the inverse filter unit 13, the power calculation unit 14, the background noise likeness calculation unit 15, the estimated background noise level update unit 16, and the estimated noise spectrum update unit 17 are used to calculate the background noise likelihood. The configuration is not limited to this configuration as long as it evaluates the likelihood.

  According to the first embodiment, by performing a predetermined signal processing process on an input signal (decoded voice), a processed signal (deformed decoding) in which a degradation component included in the input signal is not subjectively noticed Voice), and the addition weight of the input signal and the processed signal is controlled by a predetermined evaluation value (likelihood of background noise). Therefore, the ratio of the processed signal is increased centering on a section containing many degraded components. This has the effect of improving subjective quality.

  In addition, since the signal processing is performed in the spectral domain, it is possible to perform a process of suppressing a fine degradation component in the spectral domain, and there is an effect that the subjective quality can be further improved.

  In addition, since the smoothing process of the amplitude spectrum component and the disturbance imparting process of the phase spectrum component are performed as the processing, the unstable fluctuation of the amplitude spectrum component caused by quantization noise or the like can be suppressed well. Furthermore, for quantization noise, which has a unique mutual relationship between phase components and is often perceived as characteristic degradation, it is possible to disturb the relationship between phase components, thereby improving the subjective quality. There is.

  Further, the conventional binary section determination of either the speech section or the background noise section is abolished, a continuous amount called background noise likeness is calculated, and the weighted addition coefficient of the decoded speech and the modified decoded speech is continuously calculated based on this. Is controlled, so that there is an effect that quality deterioration due to a section determination error can be avoided.

  Further, when the quantization noise or the degraded sound in the voice section is large, the degraded sound can be made hard to be heard by adding the modified decoded voice even in the section that is surely known as the voice section. .

  In addition, since the output speech is generated by processing the decoded speech that contains a lot of background noise information, stable quality improvement that does not depend much on the noise type or spectrum shape while maintaining the characteristics of the actual background noise An effect is obtained, and an improvement effect is obtained even for a degradation component due to excitation coding or the like.

  Further, since the processing is performed using the decoded voice up to the present time, a particularly large delay time is not required, and there is an effect that a delay other than the processing time can be eliminated depending on the method of adding the decoded voice and the modified decoded voice. When raising the level of the modified decoded voice, the level of the decoded voice is reduced, so it is not necessary to superimpose large pseudo noise to make the quantization noise inaudible as in the past. Thus, it is possible to make the background noise level small or even large. Needless to say, since the processing is closed in the audio decoding device or the signal processing unit, it is not necessary to add new transmission information as in the related art.

  Further, in the first embodiment, the audio decoding unit and the signal processing unit are clearly separated, and information exchange between the two is small. Therefore, the audio decoding unit is introduced into various audio decoding devices including the existing one. It is easy.

Embodiment 2 FIG.
FIG. 4 shows a part of the configuration of a sound signal processing apparatus to which the sound signal processing method according to the present embodiment is applied in combination with the noise suppression method. In the figure, 36 is an input signal, 8 is a Fourier transform unit, 19 is a noise suppression unit, 39 is a spectrum transforming unit, 12 is a signal evaluation unit, 18 is a weighted addition unit, 11 is an inverse Fourier transform unit, and 40 is an output signal. is there. The spectrum deforming section 39 includes an amplitude smoothing section 9 and a phase disturbance section 10.
The operation will be described below with reference to the drawings.

  First, the input signal 36 is input to the Fourier transform unit 8 and the signal evaluation unit 12.

  The Fourier transform unit 8 performs windowing on a signal obtained by combining the input signal 36 of the input current frame and the latest portion of the input signal 36 of the previous frame as necessary, and performs Fourier transform on the signal after the windowing. By performing the conversion process, a spectrum component for each frequency is calculated, and this is output to the noise suppression unit 19. The Fourier transform process and the windowing process are the same as in the first embodiment.

  The noise suppression unit 19 subtracts the estimated noise spectrum stored inside the noise suppression unit 19 from the spectrum component for each frequency input from the Fourier transform unit 8, and weights and adds the obtained result as the noise suppression spectrum 37. The signal is output to the unit 18 and the amplitude smoothing unit 9 in the spectrum transformation unit 39. This is a process corresponding to a main part of a so-called spectral subtraction process. Then, the noise suppression unit 19 determines whether or not the noise is in the background noise interval. If the noise is in the background noise interval, the noise suppression unit 19 converts the internal estimated noise spectrum using the spectral component for each frequency input from the Fourier transform unit 8. Update. The determination as to whether or not it is in the background noise section can be simplified by making use of the output result of the signal evaluation unit 12 described later.

  The amplitude smoothing unit 9 in the spectrum transformation unit 39 performs a smoothing process on the amplitude component of the noise suppression spectrum 37 input from the noise suppression unit 19, and sends the smoothed noise suppression spectrum to the phase disturbance unit 10. Output. Regardless of the frequency axis direction or the time axis direction used here, the effect of suppressing the degraded sound generated by the noise suppression unit can be obtained. As a specific smoothing method, a method similar to that in Embodiment 1 can be used.

  The phase disturbance unit 10 in the spectrum transformation unit 39 disturbs the phase component of the smoothed noise suppression spectrum input from the amplitude smoothing unit 9, and adds the weighted spectrum as the modified noise suppression spectrum 38. Output to the unit 18. A method similar to that of the first embodiment can be used for giving a disturbance to each phase component.

  The signal evaluation unit 12 analyzes the input signal 36 to calculate the likelihood of the background noise, and outputs this as the addition control value 35 to the weighting addition unit 18. The configuration and each process in the signal evaluation unit 12 can be the same as those in the first embodiment.

    The weighted addition unit 18 converts the noise suppression spectrum 37 input from the noise suppression unit 19 and the modified noise suppression spectrum 38 input from the spectrum modification unit 39 based on the addition control value 35 input from the signal evaluation unit 12. The weighted and added spectrum is output to the inverse Fourier transform unit 11. As in the first embodiment, as the operation of the weighted addition control method, as the addition control value 35 increases (the likelihood of background noise increases), the weight for the noise suppression spectrum 37 decreases, and the weight for the modified noise suppression spectrum 38 decreases. Great control. Conversely, as the addition control value 35 becomes smaller (the lower the likelihood of the background noise), the weight for the noise suppression spectrum 37 is increased, and the weight for the modified noise suppression spectrum 38 is controlled to be smaller.

  Then, as the last process, the inverse Fourier transform unit 11 performs an inverse Fourier transform process on the spectrum input from the weighted addition unit 18 to return the spectrum to the signal region, thereby forming a smooth connection with the previous and next frames. Are connected while performing windowing for the purpose, and an obtained signal is output as an output signal 40. The windowing and connection process for connection are the same as in the first embodiment.

  According to the second embodiment, by performing a predetermined processing on the spectrum degraded by the noise suppression processing or the like, a processed spectrum (deformed noise suppression spectrum) in which the degraded component is not subjectively noticed can be obtained. Since the added weight of the spectrum before processing and the processed spectrum is controlled by a predetermined evaluation value (likelihood of background noise), a section (background noise) containing many degraded components and leading to a decrease in subjective quality is generated. The effect of improving the subjective quality is to increase the ratio of the processing spectrum around the section).

  Further, since the weighted addition is performed in the spectral domain, the Fourier transform and the inverse Fourier transform for the processing are not required as compared with the first embodiment, and the processing is simplified. Note that the Fourier transform unit 8 and the inverse Fourier transform 11 in the second embodiment are originally necessary configurations for the noise suppressing unit 19.

  In addition, since the smoothing process of the amplitude spectrum component and the disturbance imparting process of the phase spectrum component are performed as the processing, the unstable fluctuation of the amplitude spectrum component caused by quantization noise or the like can be suppressed well. Furthermore, it is possible to disturb the relationship between the phase components with respect to the quantization noise and the deteriorated components, which often have a characteristic mutual relationship between the phase components and are felt as characteristic degradation, thereby improving the subjective quality. There is an effect that can be improved.

  Also, instead of a binary section determination as to whether or not the section is a background noise section, a continuous amount called background noise likeness is calculated and the weighted addition coefficient is continuously controlled based on the calculated amount. This has the effect of avoiding quality degradation due to

  When the degraded sound is large outside the background noise section, the weighted addition as shown in FIG. 2C is performed, so that the modified noise suppression spectrum is added even in the section that is surely known to be outside the background noise section. This has the effect of making it difficult to hear the degraded sound.

  Further, since the modified noise suppression spectrum is generated by directly performing a simple process on the noise suppression spectrum, there is an effect that a stable quality improvement effect which is not so dependent on the noise type and the spectrum shape is obtained.

  In addition, since the processing is performed using the noise suppression spectrum up to the present time, in addition to the delay time of the noise suppression unit 19, there is a feature that a large delay time is not required. When increasing the level of addition of the deformed noise suppression spectrum, the level of addition of the original noise suppression spectrum is reduced, so it is not necessary to superimpose relatively large noise to prevent quantization noise from being heard. There is an effect that can be reduced. Also, needless to say, even when this processing is used as a pre-processing of the audio coding processing, the processing is closed in the coding unit, so that addition of new transmission information as in the related art is unnecessary. .

Embodiment 3 FIG.
FIG. 5 in which the same reference numerals are assigned to the parts corresponding to those in FIG. 1 shows the overall configuration of a speech decoding apparatus to which the sound signal processing method according to the present embodiment is applied. This is a deformation strength control unit that outputs information to be performed. The deformation intensity control unit 20 includes an auditory weighting unit 21, a Fourier transform unit 22, a level determination unit 23, a continuity determination unit 24, and a deformation intensity calculation unit 25.

  The operation will be described below with reference to the drawings.

  The decoded speech 5 output from the speech decoding unit 4 is input to the signal transformation unit 7, the transformation strength control unit 20, the signal evaluation unit 12, and the weighted addition unit 18 in the signal processing unit 2.

    The perceptual weighting unit 21 in the deformation intensity control unit 20 performs perceptual weighting processing on the decoded voice 5 input from the voice decoding unit 4 and outputs the obtained perceptual weighted voice to the Fourier transform unit 22. Here, as the auditory weighting processing, the same processing as that used in the audio encoding processing (which forms a pair with the audio decoding processing performed by the audio decoding unit 4) is performed.

  Perceptual weighting processing often used in coding processing such as CELP is to analyze a speech to be coded, calculate a linear prediction coefficient (LPC), and perform constant multiplication to obtain two modified LPCs. An ARMA filter using the two modified LPCs as filter coefficients is configured, and auditory weighting is performed by a filtering process using these filters. In order to perform the same auditory weighting on the decoded speech 5 as in the encoding process, an LPC obtained by decoding the received speech code 3 or an LPC calculated by re-analyzing the decoded speech 5 is used as a starting point. The two modified LPCs may be obtained and used to form an auditory weighting filter.

  In an encoding process such as CELP, encoding is performed so as to minimize distortion on the audio after the hearing weighting. In the audio after the hearing weighting, a spectral component having a large amplitude has a small amount of superposition of quantization noise. ,It turns out that. Therefore, if a speech close to the auditory weighting speech at the time of encoding can be generated in the decoding unit 1, it is useful as control information of the deformation intensity in the signal deformation unit 7.

  If the speech decoding process in the speech decoding unit 4 includes processing such as a spectrum post-filter (which is mostly included in the case of CELP), the decoded speech 5 is first processed from the decoded speech 5 originally. The audio is subjected to auditory weighting by generating a sound from which the influence of the processing such as a post filter has been removed or extracting the sound immediately before the processing from the sound decoding unit 4. A sound close to the auditory weighting sound at the time is obtained. However, when the main purpose is to improve the quality of the background noise section, the influence of the processing such as the spectrum post filter in this section is small, and there is no great difference in the effect even if the influence is not removed. The third embodiment has a configuration in which the influence of processing such as a spectral post filter is not removed.

  Needless to say, the auditory weighting unit 21 is not required when the auditory weighting is not performed in the encoding process or when the effect is small and can be ignored. In that case, the output of the Fourier transform unit 8 in the signal transforming unit 7 may be provided to the level determining unit 23 and the continuity determining unit 24, which will be described later, so that the Fourier transform unit 22 is not required.

  Furthermore, since there is a method that provides an effect similar to auditory weighting in the spectral domain, such as non-linear amplitude conversion processing, if an error from the auditory weighting method used in the encoding processing can be ignored, signal deformation is performed. The output of the Fourier transform unit 8 in the unit 7 is used as an input to the auditory weighting unit 21. The auditory weighting unit 21 performs auditory weighting on this input in the spectral domain, and the Fourier transform unit 22 is omitted. It is also possible to output an auditory weighted spectrum to the level judging unit 23 and the continuity judging unit 24 that perform the judgment.

  The Fourier transform unit 22 in the deformation intensity control unit 20 performs windowing on a signal obtained by combining the auditory weighted sound input from the auditory weighting unit 21 and the latest part of the auditory weighted audio of the previous frame as necessary, By performing a Fourier transform process on the signal after windowing, a spectrum component for each frequency is calculated, and this is output to the level determination unit 23 and the continuity determination unit 24 as an auditory weighting spectrum. The Fourier transform process and the windowing process are the same as the Fourier transform unit 8 of the first embodiment.

  The level determination unit 23 calculates a first deformation intensity for each frequency based on the magnitude of each amplitude component of the auditory weighting spectrum input from the Fourier transform unit 22, and calculates the first deformation intensity for each frequency. Output to Since the smaller the value of each amplitude component of the auditory weighting spectrum is, the larger the ratio of the quantization noise is, the first deformation strength may be increased. In the simplest case, the average value of all the amplitude components is obtained, and a predetermined threshold value Th is added to the average value. In this case, the first deformation strength may be set to 1. FIG. 6 shows the relationship between the auditory weighting spectrum and the first deformation intensity when this threshold Th is used. The method of calculating the first deformation strength is not limited to this.

  The continuity determining unit 24 evaluates the continuity of each amplitude component or each phase component of the auditory weighting spectrum input from the Fourier transform unit 22 in the time direction, and, based on the evaluation result, a second continuity for each frequency. The deformation strength is calculated, and this is output to the deformation strength calculation unit 25. It is said that good encoding was performed for the frequency component having low continuity in the time direction of the amplitude component of the auditory weighting spectrum and low continuity of the phase component (after compensating the rotation of the phase due to the time transition between frames). Is hard to imagine, so the second deformation strength is increased. For the calculation of the second deformation strength, a method of giving 0 or 1 by a determination using a predetermined threshold value can be used in the simplest manner.

  The deformation strength calculation unit 25 calculates the final deformation strength for each frequency based on the first deformation strength input from the level determination unit 23 and the second deformation strength input from the continuity determination unit 24. The calculated value is output to the amplitude smoothing unit 9 and the phase disturbance unit 10 in the signal transformation unit 7. As the final deformation strength, a minimum value, a weighted average value, a maximum value, and the like of the first deformation strength and the second deformation strength can be used. This is the end of the description of the operation of the deformation strength control unit 20 newly added in the third embodiment.

  Next, a description will be given of components whose operations are changed with the addition of the deformation strength control unit 20.

  The amplitude smoothing unit 9 performs a smoothing process on the amplitude component of the spectrum for each frequency input from the Fourier transform unit 8 according to the deformation intensity input from the deformation intensity control unit 20, and converts the smoothed spectrum. Output to the phase disturbance unit 10. It should be noted that the control is performed so that the smoothing is enhanced as the frequency component has a higher deformation intensity. The simplest method of controlling the strength of the smoothing strength is to perform smoothing only when the input deformation strength is large. Other methods of enhancing the smoothing include reducing the smoothing coefficient α in the smoothing equation described in the first embodiment, or changing the spectrum after the fixed smoothing and the spectrum before the smoothing. Are weighted and added to generate a final spectrum, and various methods such as reducing the weight of the spectrum before smoothing can be used.

  The phase disturbance unit 10 disturbs the phase component of the smoothed spectrum input from the amplitude smoothing unit 9 according to the deformation intensity input from the deformation intensity control unit 20, and converts the disturbed spectrum into an inverse Fourier transform unit 11 is output. It should be noted that the control is performed such that the greater the frequency component having the higher deformation intensity, the greater the phase disturbance. The simplest method of controlling the magnitude of the disturbance is to apply the disturbance only when the input deformation intensity is high. In addition to the above, various methods for controlling the disturbance, such as increasing or decreasing the range of the phase angle generated by random numbers, can be used.

  The other components are the same as those in the first embodiment, and a description thereof will not be repeated.

  Here, the output results of both the level determination unit 23 and the continuity determination unit 24 are used, but a configuration is possible in which only one is used and the other is omitted. Further, a configuration in which only one of the amplitude smoothing unit 9 and the phase disturbance unit 10 is controlled by the deformation intensity may be employed.

  According to the third embodiment, the magnitude of the amplitude of each frequency component of the input signal (decoded speech) or the input signal (decoded speech) weighted by the auditory sense, and the magnitude of the continuity of the amplitude and phase of each frequency Is used to control the deformation intensity at the time of generating a processed signal (deformed decoded voice) for each frequency. In addition to the effects of the first embodiment, quantization is performed because the amplitude spectrum component is small. The noise and degraded components are dominant, and the continuity of the spectral components is low, so the quantization noise and degraded components are likely to be increased. Eliminates processing of good components with few components.Effects of subjectively suppressing quantization noise and degraded components while maintaining relatively good characteristics of input signals and actual background noise, thereby improving subjective quality. There .

Embodiment 4 FIG.
FIG. 7 in which parts corresponding to those in FIG. 5 are assigned the same reference numerals shows the overall configuration of a speech decoding apparatus to which the sound signal processing method according to the present embodiment is applied. In FIG. 7, reference numeral 41 denotes an addition control value dividing unit. 5, the configuration of the signal transformation unit 7 is changed to a Fourier transform unit 8, a spectrum transformation unit 39, and an inverse Fourier transform unit 11.

  The operation will be described below with reference to the drawings.

  The decoded voice 5 output from the voice decoding unit 4 is input to the Fourier transform unit 8, the deformation intensity control unit 20, and the signal evaluation unit 12 in the signal processing unit 2.

  The Fourier transform unit 8 performs windowing on the signal obtained by combining the input decoded voice 5 of the current frame and the latest part of the decoded voice 5 of the previous frame as necessary, as in the second embodiment. A spectrum component for each frequency is calculated by performing a Fourier transform process on the signal after windowing, and this is output as a decoded speech spectrum 43 to the weighted addition unit 18 and the amplitude smoothing unit 9 in the spectrum transformation unit 39. I do.

  The spectrum deforming unit 39 performs the processing of the amplitude smoothing unit 9 and the phase disturbance unit 10 on the input decoded speech spectrum 43 in the same manner as in the second embodiment, and converts the obtained spectrum into the modified decoded speech. The spectrum 44 is output to the weighted addition unit 18.

    In the deformation intensity control unit 20, as in the third embodiment, the auditory weighting unit 21, the Fourier transform unit 22, the level determination unit 23, the continuity determination unit 24, the deformation intensity calculation The processing of the unit 25 is sequentially performed, and the obtained deformation intensity for each frequency is output to the addition control value dividing unit 41.

  Note that, as in the third embodiment, when the auditory weighting is not performed in the encoding process or when the effect is small, the auditory weighting unit 21 and the Fourier transform unit 22 become unnecessary. In that case, the output of the Fourier transform unit 8 may be provided to the level determination unit 23 and the continuity determination unit 24.

  The output of the Fourier transform unit 8 is used as an input to the auditory weighting unit 21. The auditory weighting unit 21 performs auditory weighting on this input in the spectral domain. It is also possible to configure so that the auditory weighted spectrum is output to the determination unit 23 and the continuity determination unit 24. With such a configuration, an effect of simplifying the processing can be obtained.

  As in the first embodiment, the signal evaluation unit 12 obtains the likelihood of background noise from the input decoded speech 5 and outputs this as an addition control value 35 to the addition control value division unit 41.

  The newly added addition control value division unit 41 uses the deformation intensity for each frequency input from the deformation intensity control unit 20 and the addition control value 35 input from the signal evaluation unit 12 to perform addition control for each frequency. A value 42 is generated and output to the weighted addition unit 18. For a frequency having a high deformation intensity, the value of the addition control value 42 of the frequency is controlled so that the weight of the decoded voice spectrum 43 in the weighted addition section 18 is weakened and the weight of the deformed decoded voice spectrum 44 is strong. Conversely, for a frequency having a low deformation intensity, the value of the addition control value 42 of that frequency is controlled so that the weight of the decoded voice spectrum 43 in the weighted addition unit 18 is increased and the weight of the deformed decoded voice spectrum 44 is reduced. That is, for a frequency having a high deformation intensity, the likelihood of the background noise is high. Therefore, the addition control value 42 of the frequency is increased, and in the opposite case, it is decreased.

  The weighted adder 18 converts the decoded speech spectrum 43 input from the Fourier transformer 8 and the deformation input from the spectrum converter 39 based on the addition control value 42 for each frequency input from the addition control value divider 41. The decoded speech spectrum 44 is weighted and added, and the obtained spectrum is output to the inverse Fourier transform unit 11. The operation of the weighted addition control method is similar to that described with reference to FIG. 2, in which a weight for the decoded speech spectrum 43 is assigned to a frequency component having a large addition control value 42 for each frequency (high background noise likeness). The weight for the modified decoded speech spectrum 44 is controlled to be large and small. Conversely, for a frequency component in which the addition control value 42 for each frequency is small (the likelihood of background noise is low), the weight for the decoded speech spectrum 43 is increased, and the weight for the modified decoded speech spectrum 44 is controlled to be small.

  Then, as the last process, the inverse Fourier transform unit 11 performs an inverse Fourier transform process on the spectrum input from the weighted addition unit 18 in the same manner as in Embodiment 2, thereby returning the spectrum to the signal domain. The connection is performed while performing windowing for smooth connection with the preceding and succeeding frames, and the obtained signal is output as output sound 6.

  Note that the addition control value division unit 41 is omitted, the output of the signal evaluation unit 12 is given to the weighted addition unit 18, and the deformation intensity output from the deformation intensity control unit 20 is sent to the amplitude smoothing unit 9 and the phase disturbance unit 10. A configuration for providing is also possible. Such a configuration corresponds to a configuration in which the weighted addition process in the configuration of the third embodiment is performed in the spectral domain.

Further, similarly to the case of the third embodiment, a configuration is possible in which only one of the level determination unit 23 and the continuity determination unit 24 is used, and the other is omitted.
According to the fourth embodiment, the magnitude of the amplitude of each frequency component of the input signal (decoded speech) or the input signal (decoded speech) weighted by auditory sense, and the magnitude of the continuity of the amplitude and phase of each frequency , The weighted addition of the spectrum of the input signal (decoded speech spectrum) and the processed spectrum (deformed decoded speech spectrum) are controlled independently for each frequency component. In addition to the effects of the first embodiment, The components in which the quantization noise and the degraded components are dominant because the amplitude spectrum components are small, and the components in which the quantization noise and the degraded components tend to be increased due to the low continuity of the spectral components. By increasing the weight of the processing spectrum, the weight of the processing spectrum will not be strengthened even to a good component with less quantization noise and degraded components. While leaving sex relatively good can subjectively suppressed quantization noise or the degraded component, there is an effect of improving the subjective quality.

  Compared with the third embodiment, the transformation processing for each frequency, namely, smoothing and disturbance, is changed to the transformation processing for each frequency, which has the effect of simplifying the processing.

Embodiment 5 FIG.
FIG. 8 in which the same reference numerals are assigned to parts corresponding to those in FIG. 5 shows the entire configuration of a speech decoding apparatus to which the sound signal processing method according to the present embodiment is applied. In the figure, reference numeral 26 denotes background noise likeness (addition control value 35). Is a variability determination unit that determines the variability in the time direction of.

  The operation will be described below with reference to the drawings.

  The decoded speech 5 output from the speech decoding unit 4 is input to the signal transformation unit 7, the transformation strength control unit 20, the signal evaluation unit 12, and the weighted addition unit 18 in the signal processing unit 2. The signal evaluation unit 12 evaluates the likelihood of background noise with respect to the input decoded speech 5 and outputs the evaluation result as an addition control value 35 to the variability determination unit 26 and the weighted addition unit 18.

The variability determination unit 26 compares the addition control value 35 input from the signal evaluation unit 12 with a past addition control value 35 stored therein, and determines whether the variability of the value in the time direction is high. Is determined, and a third deformation strength is calculated based on the determination result, and this is output to the deformation strength calculation unit 25 in the deformation strength control unit 20. Then, the past addition control value 35 stored inside is updated using the input addition control value 35.
If the time direction variability of a parameter representing the characteristics of a frame (or a subframe) such as the addition control value 35 is high, the spectrum of the decoded speech 5 often changes greatly in the time direction, which is more than necessary. When strong amplitude smoothing or phase disturbance is applied, an unnatural reverberation is generated. Therefore, the third deformation strength is set such that when the variability in the time direction of the addition control value 35 is high, the smoothing in the amplitude smoothing unit 9 and the disturbance application in the phase disturbance unit 10 are weak. Note that the same effect can be obtained by using parameters other than the addition control value 35, such as the power of the decoded speech and the spectrum envelope parameter, as long as the parameters represent the characteristics of the frame (or the subframe).

  The simplest method of determining the variability is to compare the absolute value of the difference from the addition control value 35 of the previous frame with a predetermined threshold, and determine that the variability is high if the absolute value exceeds the threshold. . Alternatively, the absolute value of the difference between the addition control value 35 of the previous frame and the two frames before the previous frame may be calculated, and the determination may be made based on whether one of them exceeds a predetermined threshold. When the signal evaluation unit 12 calculates the addition control value 35 for each subframe, the signal evaluation unit 12 calculates the absolute value of the difference between the addition control values 35 between all the subframes in the current frame or, if necessary, in the previous frame. In this case, it can be determined whether or not any of them exceeds a predetermined threshold. Then, as a specific processing example, the third deformation intensity is set to 0 when the value exceeds the threshold value, and the third deformation intensity is set to 1 when the value is lower than the threshold value.

  In the deformation strength control unit 20, the same processing as in the third embodiment is performed on the input decoded speech 5 up to the auditory weighting unit 21, the Fourier transform unit 22, the level determination unit 23, and the continuity determination unit 24. Do.

  Then, in the deformation strength calculation unit 25, the first deformation strength input from the level determination unit 23, the second deformation strength input from the continuity determination unit 24, and the third deformation intensity input from the variability determination unit 26 Based on the deformation intensity, a final deformation intensity for each frequency is calculated and output to the amplitude smoothing unit 9 and the phase disturbance unit 10 in the signal deformation unit 7. As a method of calculating the final deformation strength, the third deformation strength is given as a constant value for all frequencies, and the third deformation strength, first deformation strength, A method of obtaining a minimum value, a weighted average value, a maximum value, and the like of the second deformation strength to obtain a final deformation strength can be used.

  The subsequent operations of the signal transformation unit 7 and the weighted addition unit 18 are the same as in the third embodiment, and a description thereof will be omitted.

  Here, although the output results of both the level determination unit 23 and the continuity determination unit 24 are used, a configuration in which only one of them is used or both are not used is also possible. Further, a configuration in which only one of the amplitude smoothing unit 9 and the phase disturbance unit 10 is controlled by the deformation intensity or a configuration in which only one of the third deformation intensity is controlled may be used.

  According to the fifth embodiment, in addition to the configuration of the third embodiment, the smoothing strength or the disturbance imparting strength is determined by changing the temporal variability (variability between frames or subframes) of a predetermined evaluation value (likelihood of background noise). ), It is possible to suppress unnecessarily strong processing in a section where the characteristics of the input signal (decoded voice) fluctuate, in addition to the effects of the third embodiment. This has the effect of preventing the occurrence of echo.

Embodiment 6 FIG.
FIG. 9 in which parts corresponding to those in FIG. 5 are assigned the same reference numerals shows the overall configuration of a speech decoding apparatus to which the sound signal processing method according to the present embodiment is applied. In the figure, reference numeral 27 denotes a friction noise likeness evaluation unit, 31 denotes a background noise likeness evaluation unit, and 45 denotes an addition control value calculation unit. The fricative sound likeness evaluation unit 27 includes a low frequency cut filter 28, a zero-crossing number counting unit 29, and a fricative sound likeness calculating unit 30. The background noise likeness evaluation unit 31 has the same configuration as the signal evaluation unit 12 in FIG. 5, and includes an inverse filter unit 13, a power calculation unit 14, a background noise likeness calculation unit 15, an estimated noise power update unit 16, and an estimated noise spectrum update unit. 17. The signal evaluation unit 12 is different from the case of FIG. 5 and includes a friction noise likeness evaluation unit 27, a background noise likeness evaluation unit 31, and an addition control value calculation unit 45.

  The operation will be described below with reference to the drawings.

  The decoded speech 5 output from the speech decoding unit 4 is output to the signal transformation unit 7, the deformation intensity control unit 20 in the signal processing unit 2, the fricative noise likeness evaluation unit 27 and the background noise likeness evaluation unit 31 in the signal evaluation unit 12, and It is input to the weighted addition unit 18.

  Like the signal evaluation unit 12 in the third embodiment, the background noise likeness evaluation unit 31 in the signal evaluation unit 12 applies the inverse filter unit 13, the power calculation unit 14, and the background noise likeness to the input decoded speech 5. The processing of the calculation unit 15 is performed, and the obtained background noise likeness 46 is output to the addition control value calculation unit 45. Further, the processing of the estimated noise power updating unit 16 and the estimated noise spectrum updating unit 17 is performed to update the estimated noise power and the estimated noise spectrum stored in each of them.

  The low-frequency cut filter 28 in the fricative soundness evaluation section 27 performs low-frequency cut filtering processing on the input decoded voice 5 to suppress low frequency components, and outputs the filtered decoded voice to the zero-crossing number counting section 29. Output. The purpose of the low-frequency cut filtering processing is to prevent a DC component or a low-frequency component included in the decoded voice from becoming an offset and reducing the count result of the zero-crossing number counting unit 29 described later. Therefore, simply, the average value of the decoded speech 5 in the frame may be calculated and subtracted from each sample of the decoded speech 5.

  The number-of-zero-crossings counting unit 29 analyzes the voice input from the low-frequency cut filter 28, counts the number of included zero-crossings, and outputs the obtained number of zero-crossings to the fricative likelihood calculating unit 30. As a method of counting the number of zero crossings, a method of comparing the positive and negative of adjacent samples, counting them if they are not the same, counting them as zero, and taking the product of the values of the adjacent samples and calculating the result as negative or zero If there is, there is a method of counting as crossing zero.

  The fricative sound likeness calculating section 30 compares the number of zero crossings input from the zero crossing number counting section 29 with a predetermined threshold value, obtains a fricative likelihood 47 based on the comparison result, and adds this to the addition control value calculating section 45. Output to For example, when the number of zero crossings is larger than the threshold value, it is determined that the sound is a fricative sound, and the likelihood of the fricative sound is set to 1. Conversely, if the number of zero crossings is smaller than the threshold value, it is determined that the sound is not like a fricative sound, and the likelihood of fricative sound is set to 0. In addition, two or more threshold values may be provided to set the likelihood of frictional noise in a stepwise manner, or a predetermined function may be prepared to calculate the likelihood of continuous frictional noise from the number of zero crossings. .

  Note that the configuration in the fricative soundness evaluation section 27 is merely an example, and the evaluation is performed based on the analysis result of the spectrum tilt, the evaluation is performed based on the power and the stationarity of the spectrum, or the zero-crossing is performed. The evaluation may be performed by combining a plurality of parameters including the number.

  The addition control value calculation unit 45 calculates an addition control value 35 based on the background noise likelihood 46 input from the background noise likeness evaluation unit 31 and the fricative sound likeness 47 input from the fricative sound likeness evaluation unit 27. Output to the weighted addition unit 18. Since the quantization noise often becomes hard to hear in both the case of the background noise and the case of the fricative sound, the addition control value 35 is calculated by appropriately weighting and adding the background noise likeness 46 and the fricative sound likeness 47. do it.

    Subsequent operations of the signal transformation unit 7, the transformation intensity control unit 20, and the weighted addition unit 18 are the same as those in the third embodiment, and a description thereof will be omitted.

  According to the sixth embodiment, when the likelihood of background noise and fricative noise of an input signal (decoded voice) is high, a processed signal (modified decoded voice) is output more instead of the input signal (decoded voice). Therefore, in addition to the effects of the third embodiment, emphasis processing is applied to a friction sound section in which quantization noise and a large amount of degradation components tend to occur, and a section other than the friction sound is appropriately processed in that section. Since processing (no processing, low-level processing, etc.) is selected, there is an effect that the subjective quality can be improved. In addition, when a portion where quantization noise and a large amount of degradation components tend to occur can be specified to some extent other than the frictional sound, it is possible to evaluate the likelihood and reflect it in the addition control value. With such a configuration, since large quantization noise and degraded components can be suppressed one by one, there is an effect that the subjective quality can be further improved.

  Of course, a configuration in which the background noise likeness evaluation unit is deleted is also possible.

Embodiment 7 FIG.
FIG. 10 in which the same reference numerals are assigned to parts corresponding to those in FIG. 1 shows the entire configuration of a speech decoding device to which the signal processing method according to the present embodiment is applied, and 32 in the figure denotes a post-filter unit.

  The operation will be described below with reference to the drawings.

  First, the speech code 3 is input to the speech decoding unit 4 in the speech decoding device 1.

  The audio decoding unit 4 performs a decoding process on the input audio code 3 and outputs the obtained decoded audio 5 to the post-filter unit 32, the signal transformation unit 7, and the signal evaluation unit 12.

  The post-filter unit 32 performs a spectrum enhancement process, a pitch periodicity enhancement process, and the like on the input decoded speech 5, and outputs the obtained result to the weighted addition unit 18 as a post-filter decoded speech 48. This post-filter process is generally used as a post-process of the CELP decoding process, and is introduced for the purpose of suppressing quantization noise generated by encoding and decoding. Since the quantization noise is much contained in the portion where the spectrum intensity is weak, the amplitude of this component is suppressed. In some cases, the pitch periodicity enhancement process is not performed, and only the spectrum enhancement process is performed.

  In the first embodiment and the third to sixth embodiments, a description has been given of a case in which this post-filter processing is applicable to both those included in the audio decoding unit 4 and those not present. 7, all or a part of the post-filter processing is made independent as the post-filter unit 32 from the post-filter processing included in the audio decoding unit 4.

  The signal transformation unit 7 performs the processing of the Fourier transform unit 8, the amplitude smoothing unit 9, the phase disturbance unit 10, and the inverse Fourier transform unit 11 on the input decoded speech 5 as in the first embodiment. The obtained modified decoded speech 34 is output to the weighted addition unit 18.

  As in the first embodiment, the signal evaluation unit 12 evaluates the likelihood of background noise with respect to the input decoded speech 5 and outputs the evaluation result to the weighted addition unit 18 as an addition control value 35.

  Then, as the last process, the weighted addition unit 18 performs the post-filter decoding voice input from the post-filter unit 32 based on the addition control value 35 input from the signal evaluation unit 12 as in the first embodiment. 48 and the modified decoded speech 34 input from the signal transformation unit 7 are weighted and added, and the obtained output speech 6 is output.

  According to the seventh embodiment, a modified decoded speech is generated based on the decoded speech before processing by the post filter, and the decoded speech before processing by the post filter is analyzed to determine the likelihood of background noise. Since the weight at the time of adding the post-filter decoded speech and the modified decoded speech is controlled, in addition to the effect of the first embodiment, a modified decoded speech that does not include the modification of the decoded speech by the post filter can be generated. Since highly accurate addition weight control can be performed based on the high-accuracy background noise calculated without being affected by the deformation of the decoded speech by the filter, the subjective quality is further improved.

  In the background noise section, even the degraded sound is often emphasized by the post filter, making it difficult to hear. Becomes smaller. Also, post-filter processing has multiple modes, and if the processing is frequently switched, there is a high risk that the switching will affect the evaluation of the likelihood of background noise. A more stable evaluation result can be obtained by evaluating the likelihood of the background noise.

  In the configuration of the third embodiment, when the post-filter unit is separated in the same manner as in the seventh embodiment, the output result of the auditory weighting unit 21 in FIG. As the sound approaches the voice, the accuracy of specifying a component having a large amount of quantization noise increases, better deformation intensity control can be obtained, and the effect of further improving the subjective quality can be obtained.

  Further, in the configuration of the sixth embodiment, when the post-filter unit is separated in the same manner as in the seventh embodiment, the evaluation accuracy in the frictional sound likeness evaluation unit 27 in FIG. 9 is increased, and the subjective quality is further improved. The effect is obtained.

  The configuration in which the post filter section is not separated has a smaller connection with the audio decoding section (including the post filter) to only one point of the decoded voice than the separated configuration of the seventh embodiment. There is an advantage that it can be easily realized by a device and a program. In the seventh embodiment, there is a disadvantage that it is not easy to realize an audio decoding unit having a post filter by an independent device and a program, but it has the various effects described above.

Embodiment 8 FIG.
FIG. 11, in which parts corresponding to those in FIG. 10 are assigned the same reference numerals, shows the entire configuration of a speech decoding apparatus to which the sound signal processing method according to the present embodiment is applied. It is a spectrum parameter. The difference from FIG. 10 is that a deformation intensity control unit 20 similar to that of the third embodiment is added, and a spectrum parameter 33 is input from the speech decoding unit 4 to the signal evaluation unit 12 and the deformation intensity control unit 20. It is.

  The operation will be described below with reference to the drawings.

  First, the speech code 3 is input to the speech decoding unit 4 in the speech decoding device 1.

  The audio decoding unit 4 performs a decoding process on the input audio code 3 and outputs the obtained decoded audio 5 to the post-filter unit 32, the signal modification unit 7, the modification intensity control unit 20, and the signal evaluation unit 12. . In addition, the spectrum parameters 33 generated in the decoding process are output to the estimated noise spectrum updating unit 17 in the signal evaluation unit 12 and the auditory weighting unit 21 in the deformation intensity control unit 20. Note that, as the spectrum parameter 33, a linear prediction coefficient (LPC), a line spectrum pair (LSP), and the like are often used in general.

  The auditory weighting unit 21 in the deformation intensity control unit 20 performs an auditory weighting process on the decoded speech 5 input from the audio decoding unit 4 using the spectrum parameter 33 also input from the audio decoding unit 4 and obtains The obtained auditory weighted sound is output to the Fourier transform unit 22. As a specific process, when the spectrum parameter 33 is a linear prediction coefficient (LPC), this is used as it is, and when the spectrum parameter 33 is a parameter other than LPC, the spectrum parameter 33 is converted into LPC. Then, the LPC is multiplied by a constant to obtain two modified LPCs, an ARMA filter using the two modified LPCs as filter coefficients is formed, and auditory weighting is performed by a filtering process using this filter. In addition, it is desirable that the auditory weighting process performs the same process as that used in the audio encoding process (which forms a pair with the audio decoding process performed by the audio decoding unit 4).

  In the deformation intensity control unit 20, following the processing of the auditory weighting unit 21, the processing of the Fourier transform unit 22, the level determination unit 23, the continuity determination unit 24, and the deformation intensity calculation unit 25, as in the third embodiment. And outputs the obtained deformation intensity to the signal deformation unit 7.

  Similarly to the third embodiment, the signal transformation unit 7 processes the input decoded speech 5 and the transformation strength by the Fourier transform unit 8, the amplitude smoothing unit 9, the phase disturbance unit 10, and the inverse Fourier transform unit 11. And outputs the obtained modified decoded speech 34 to the weighted addition unit 18.

  In the signal evaluation unit 12, as in the first embodiment, the input decoded speech 5 is first subjected to the processing of the inverse filter unit 13, the power calculation unit 14, and the background noise likeness calculation unit 15, and the background noise likeness is thus obtained. And outputs the evaluation result to the weighted addition unit 18 as an addition control value 35. Further, the processing of the estimated noise power updating unit 16 is performed to update the internal estimated noise power.

The estimated noise spectrum updating unit 17 updates the estimated noise spectrum stored therein using the spectrum parameters 33 input from the speech decoding unit 4 and the background noise input from the background noise likeness calculating unit 15. . For example, when the likelihood of the input background noise is high, the update is performed by reflecting the spectrum parameter 33 in the estimated noise spectrum according to the equation shown in the first embodiment.
Subsequent operations of the post-filter unit 32 and the weighted addition unit 18 are the same as those in the seventh embodiment, and thus description thereof is omitted.

  According to the eighth embodiment, the auditory weighting process and the update of the estimated noise spectrum are performed by using the spectrum parameters generated in the speech decoding process. Therefore, the third and seventh embodiments are used. There is an effect that processing is simplified, in addition to the effect of.

  Furthermore, the same auditory weighting process as the encoding process is realized, the accuracy of specifying a component with much quantization noise is increased, better deformation intensity control is obtained, and the effect of improving the subjective quality is obtained.

  In addition, the estimation accuracy of the estimated noise spectrum used for calculating the likelihood of background noise (in the sense that it is close to the spectrum of the speech input to the speech encoding process) is improved, and the resulting stable and highly accurate background noise is considered. Based on this, it is possible to perform highly accurate addition weight control, and there is an effect that the subjective quality is improved.

  Although the post-filter unit 32 is separated from the audio decoding unit 4 in the eighth embodiment, the spectrum output from the audio decoding unit 4 as in the eighth embodiment may be used even in a non-separated configuration. The processing of the signal processing unit 2 can be performed using the parameter 33. In this case, the same effect as in the eighth embodiment can be obtained.

Embodiment 9 FIG.
In the configuration of the fourth embodiment shown in FIG. 7, the outline of the spectrum after the addition control value division unit 41 multiplies the weight for each frequency of the modified decoded speech spectrum 44 added by the weighting addition unit 18 is as follows. It is also possible to control the output deformation intensity so as to match the estimated spectral shape of the quantization noise.

  FIG. 12 is a schematic diagram showing an example of the spectrum after multiplying the decoded speech spectrum 43 and the modified decoded speech spectrum 44 by the weight for each frequency in this case.

  On the decoded speech spectrum 43, quantization noise having a spectrum shape depending on the encoding method is superimposed. In the CELP speech coding method, a code search is performed so as to minimize distortion in the speech after the auditory weighting process. For this reason, the quantization noise has a flat spectrum shape in the speech after the hearing weighting process, and the final quantization noise spectrum shape has a spectrum shape of the inverse characteristic of the hearing weighting process. become. Therefore, it is possible to obtain the spectrum characteristic of the auditory weighting process, obtain the spectrum shape of the inverse characteristic, and control the output of the addition control value dividing unit 41 so that the spectrum shape of the modified decoded speech spectrum matches this. is there.

  According to the ninth embodiment, the spectrum shape of the modified decoded speech component included in the final output speech 6 is made to match the approximate shape of the estimated spectrum of the quantization noise. In addition to the above, there is an effect that it is possible to make the hard-to-hear quantization noise difficult to hear in the voice section by adding the deformed decoded voice with the minimum necessary power.

Embodiment 10 FIG.
In the configuration of the first embodiment and the third to eighth embodiments, in the processing of the amplitude smoothing unit 9, processing may be performed such that the amplitude spectrum after smoothing matches the amplitude spectrum shape of the estimated quantization noise. It is possible. The calculation of the amplitude spectrum shape of the estimated quantization noise may be performed in the same manner as in the ninth embodiment.

  According to the tenth embodiment, the spectrum shape of the modified decoded speech is made to match the estimated spectrum shape of the quantization noise. Therefore, in addition to the effects of the first and third to eighth embodiments, necessary By adding the modified decoded voice with the minimum power, there is an effect that uncomfortable quantization noise in a voice section can be hardly heard.

Embodiment 11 FIG.
In the first embodiment and the third to tenth embodiments, the signal processing unit 2 is used for processing the decoded voice 5. However, only the signal processing unit 2 is extracted and the audio signal decoding unit (the audio signal encoding unit) is used. , And can be used for other signal processing, such as connecting to the subsequent stage of the noise suppression processing. However, it is necessary to change and adjust the deformation process in the signal deformation unit and the evaluation method in the signal evaluation unit according to the characteristics of the degradation component to be eliminated.

  According to the eleventh embodiment, it is possible to process a signal including a degraded component other than the decoded voice so that a subjectively undesirable component is less likely to be perceived.

Embodiment 12 FIG.
In the first to eleventh embodiments, the signal is processed by using the signal up to the current frame. However, a configuration in which the processing delay is allowed to use the signal of the next frame and thereafter is also possible.

  According to the twelfth embodiment, it is possible to refer to the signal after the next frame, so that it is possible to obtain the effect of improving the smoothing characteristics of the amplitude spectrum, improving the accuracy of continuity determination, and improving the evaluation accuracy such as noise.

Embodiment 13 FIG.
In the first, third, and fifth to twelfth embodiments, the spectral components are calculated by the Fourier transform, transformed, and returned to the signal domain by the inverse Fourier transform. Alternatively, it is also possible to perform a transformation process on each output of the band-pass filter group and reconstruct the signal by adding the signals for each band.

  According to the thirteenth embodiment, the same effect can be obtained even in a configuration not using Fourier transform.

Embodiment 14 FIG.
In the first to thirteenth embodiments, the configuration includes both the amplitude smoothing unit 9 and the phase disturbance unit 10. However, a configuration in which one of the amplitude smoothing unit 9 and the phase disturbance unit 10 is omitted is also possible. However, a configuration in which another deformed portion is introduced is also possible.

  According to the fourteenth embodiment, depending on the characteristics of the quantization noise and the degraded sound to be eliminated, there is an effect that the processing can be simplified by omitting a deformed portion having no introduction effect. In addition, by introducing an appropriate deformation unit, an effect of eliminating quantization noise and degraded sound that cannot be eliminated by the amplitude smoothing unit 9 and the phase disturbance unit 10 can be expected.

FIG. 1 is a diagram illustrating an overall configuration of a speech decoding device to which a speech decoding method according to Embodiment 1 of the present invention has been applied. FIG. 5 is a diagram illustrating a control example of weighted addition based on an addition control value in the weighted addition unit 18 according to the first embodiment of the present invention. FIG. 4 is an explanatory diagram illustrating an example of an actual shape of a cutout window in the Fourier transform unit 8, a window for connection in the inverse Fourier transform unit 11, and a time relationship with the decoded speech 5 according to the first embodiment of the present invention. FIG. 14 is a diagram illustrating a part of a configuration of a speech decoding device to which the sound signal processing method according to the second embodiment of the present invention is applied in combination with a noise suppression method. FIG. 13 is a diagram illustrating an overall configuration of a speech decoding device to which a speech decoding method according to Embodiment 3 of the present invention is applied. FIG. 13 is a diagram illustrating a relationship between an auditory weighting spectrum and a first deformation intensity according to the third embodiment of the present invention. FIG. 14 is a diagram illustrating an overall configuration of a speech decoding device to which a speech decoding method according to Embodiment 4 of the present invention is applied. FIG. 15 is a diagram illustrating an overall configuration of a speech decoding device to which a speech decoding method according to a fifth embodiment of the present invention is applied. FIG. 15 is a diagram illustrating an overall configuration of a speech decoding device to which a speech decoding method according to Embodiment 6 of the present invention is applied. FIG. 15 is a diagram illustrating an overall configuration of a speech decoding device to which a speech decoding method according to a seventh embodiment of the present invention is applied. FIG. 21 is a diagram illustrating an overall configuration of a speech decoding device to which a speech decoding method according to an eighth embodiment of the present invention is applied. FIG. 21 is a schematic diagram showing an example of a spectrum obtained by multiplying a decoded speech spectrum 43 to which Embodiment 9 of the present invention is applied and a modified decoded speech spectrum 44 by a weight for each frequency.

Claims (1)

Decoding a speech code to generate a decoded speech, a decoded speech generation step of generating predetermined information based on the speech code,
A first processed voice generation step of processing the decoded voice to generate a first processed voice,
An evaluation value calculation step of calculating a predetermined evaluation value based on the information,
A second processed voice generating step of generating a second processed voice by weighting and adding the decoded voice and the first processed voice based on the evaluation value,
An output audio step of outputting the second processed audio as an output audio.

Images