JPH08248997A - Voice band enlarging device - Google Patents

Abstract

PURPOSE: To provide a clear voice with band feeling by enlarging a band limited voice band such as a telephone voice, etc., and synthesizing a wide band voice. CONSTITUTION: This device is composed of an LPC analysis part 12 extracting a spectrum envelope information and a residual waveform from an input signal, an envelope enlargement part 13 band enlarging the extracted spectrum envelope information, a drive voice source enlargement part 14 band enlarging an extracted residual signal, an LPC synthesis part 15 generating a synthetic voice by using the spectrum envelope information obtained from the envelope enlargement part and the residual signal obtained from the drive voice source enlargement part 14, a waveform smoothing part 16 smoothing the synthetic voice on a time base, a filter part 17 filtering a frequency band to be added to the input signal from the smoothed signal, an upsampling part 11 converting the input signal to a high sampling frequency and an addition part 18 adding an output signal from the filter part 17 to the output signal from the upsampling part 11.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention expands the band from a narrow band audio signal to a wide band audio signal in an environment where only a narrow band audio signal can be obtained, thereby providing a voice with clarity and a sense of band. The present invention relates to a voice band expansion device provided.

[0002]

2. Description of the Related Art In recent years, several methods for expanding a voice band have been proposed by academic societies. Most of them focus on expanding the bandwidth of telephone voice, and the effects of each method have been announced. The conventional voice band expansion method will be described below.

FIG. 4 is a technical report of the Institute of Electronics, Information and Communication Engineers (SP93
-61) This is a method proposed by Yoshida, Abe et al. In "Reconstruction method of narrow band speech to wide band speech by codebook mapping", and is a device configuration diagram for realizing the speech band expanding method of the first conventional example. Is shown. In FIG. 4, 41 is an LPC analysis unit, 42 is a vector quantization unit, 43 is a decoding unit, 44 is a narrow band codebook, 45 is a wideband codebook, 46 is a low band restoration unit, and 47 is a first high band. A restoration unit, 48 is a second high frequency restoration unit, 49 is an addition unit, and 410 is an upsampling unit. The operation will be described below.

The input narrowband speech signal is subjected to linear prediction analysis in the LPC analysis section 41. The spectral envelope information extracted by the LPC analysis unit 41 is converted into the narrowband codebook 4
4, the vector quantizer 42 performs vector quantization. The decoding unit 43 obtains wideband spectrum envelope information stored in the wideband codebook 45 corresponding to the narrowband codebook 44 using the result of the vector quantization unit 42. In the low-frequency restoration unit 46, the LPC analysis unit 41
Using the pitch information, the power information, and the spectrum envelope information obtained from the decoding unit 43, the low frequency band is restored by LPC synthesis. Similarly, the first high frequency band restoration unit 47 restores the high frequency band based on the pitch information, the power information, and the spectrum envelope information obtained from the decoding unit 43. An adder 49 algebraically adds the outputs of the low-band restorer 46 and the first high-band restorer 47 and the input narrowband signal whose sampling frequency has been converted by the upsampling unit 410 to obtain a wideband. Get synthetic speech.

Further, in this first conventional example, a method of using the second high-frequency restoring unit 48 instead of the first high-frequency restoring unit 47 is also proposed, and the second high-frequency restoring unit 48 is proposed. Then, referring to the output of the vector quantizer 42, the high band is restored from the accumulated waveform segment data, the pitch information and the power information from the LPC analyzer 41, and the output of the low band restorer 46 The output from the upsampling unit 410 is added to the adding unit 49.
To obtain wideband synthesized speech. Other configurations are
This is the same as the method using the first high frequency band restoration unit 47.

[0006] FIG. 5 is a 1-P-
6 (1994-11), which is a method proposed by Mr. Tanaka et al. In "Examination of Pseudo Wide Band of Telephone Band Voice", showing a device configuration diagram for realizing the second conventional voice band expanding method. is there. In FIG. 5, 51 is an LPC analysis unit, 52 is an LPC inverse filter unit, 53 is a rectification unit, and 54.
Is an LPC synthesis unit, 55 is a band pass filter unit, 56 is a high pass filter unit, and 57 is an equalizer unit. The operation will be described below.

The input narrow band signal is processed by the LPC analyzer 5
1 for linear predictive analysis. Using the output obtained from the LPC analysis unit 51, the LPC inverse filter unit 52 calculates the residual signal. The residual signal is non-linearly processed by the rectifying unit 53 using half-wave rectification. Based on the output of the rectification unit 53 and the spectrum envelope information obtained from the LPC analysis unit 51, the LPC
Linear combination is performed in the combining unit 54. In the band pass filter 55, 0 to 300 is calculated from the input narrow band signal.
Filter the Hz band. In the high pass filter unit 56,
From the input narrow band signal, high frequency components in the 300 Hz to 4 kHz band are filtered. The output of the band-pass filter 55 and the output of the high-pass filter 56 are added, and the equalizer unit 57 performs frequency equalization so that the synthesized voice does not have a muffled feeling, thereby obtaining a wide-band output voice.

[0008] FIG. 6 is a collection of papers of the Acoustical Society of Japan 1-5.
9 (1994-11) "Reconstruction of wideband speech from narrowband CELP code" proposed by Tasaki et al., Showing an apparatus configuration diagram for realizing the third conventional speech band expansion method. Is. In FIG. 6, 61 is an upsampling unit, 62 is an LSP analysis unit, 63 is an adaptive sound source analysis unit, 64 is a wideband LSP estimation unit, and 65 is an LS.
A P synthesis filter unit, 66 is an adaptive sound source synthesis unit, 67 is a noise generation unit, 68 is a high-pass filter unit, 69 and 610 are multipliers, and 611 and 612 are adders. The operation will be described below.

The input narrow band signal is analyzed by the LSP analysis section 62 using the LSP analysis of the linear prediction analysis. That is, the well-known CELP (Code Exc
After it is analyzed using the idea of itedLinear Prediction), the adaptive sound source analysis unit 63 outputs the narrow band adaptive sound source gain, the adaptive sound source lag length, and the residual gain. The adaptive sound source lag length is doubled, and using this value, the adaptive sound source synthesizing unit 66
To synthesize an adaptive sound source. The output from the adaptive sound source synthesis unit 66 is multiplied by a value obtained by doubling the adaptive sound source gain in a multiplier 69. In addition, the noise generation unit 67 generates a noise component from the noise codebook, and this noise component is added to the multiplier 61.
At 0, it is multiplied by a value obtained by doubling the residual gain. A wide band sound source is generated by adding the output of the multiplier 69 and the output of the multiplier 610 by the adder 611. By reducing the order of the LSP from the LSP analysis unit 62, the wideband LSP estimation unit 64 widens the pseudo band. Adder 611
The wideband sound source output from the LSP synthesis filter unit 65 and the output from the wideband LSP estimation unit 64 are used.
Wideband speech is synthesized with. A signal obtained by extracting a high-frequency component of 3400 Hz to 7200 Hz by the high-pass filter unit 68 from the wideband sound, and up-sampling the input narrow-band signal by the up-sampling unit 61;
The output of the high-pass filter unit 68 is added by the adder 612 to obtain a wideband synthesized sound.

[0010]

However, in the above-mentioned conventional configuration, first, in the method of the first conventional example, since the sound source information is pitch and white noise, sound quality deterioration due to a pitch estimation error becomes a problem. In addition, a large amount of calculation and memory are required to estimate the envelope information.

In the second conventional method, 0 Hz to 300 Hz
Only the band of Hz is expanded, and the band of 3400 Hz or higher is not expanded, so that a band feeling cannot be obtained.

In the method of the third conventional example, a low frequency band of 300 Hz or less is not generated, and sufficient naturalness cannot be obtained. In addition, widening the spectrum envelope information by reducing the LSP order has a problem that the broadband spectrum envelope information cannot be restored with sufficient accuracy.

The present invention solves the above-mentioned conventional problems, and an object of the present invention is to provide a speech band expanding device which is clear and has a sense of band by synthesizing both low frequency components and high frequency components. To do.

[0014]

In order to achieve this object, a speech band expanding apparatus of the present invention comprises an LPC analysis section for extracting spectral information and a residual signal from an input signal using a linear prediction analysis method. An envelope expansion unit that expands the spectrum envelope information extracted by the LPC analysis unit to a desired frequency band by a linear mapping function, and a non-linear processing of the residual signal calculated by the LPC analysis unit,
Using a driving sound source expansion unit for converting into a signal having a desired frequency band characteristic, band expanded envelope information obtained by the envelope expansion unit, and band expanded sound source information obtained by the driving sound source expansion unit. LPC that synthesizes voice with LPC
A combining unit, a waveform smoothing unit that smoothes the signal sequence obtained from the LPC combining unit on the time axis, and a frequency band component to be added to the input signal from the signal obtained from the waveform smoothing unit. A filter section for filtering, an up-sampling section for outputting a signal sequence in which the input signal is converted to a sampling frequency higher than the sampling frequency of the input signal while maintaining frequency characteristics, an output signal of the filter section and the up-sampling section. And an adder that adds the output signals of the sampling unit, and is configured to obtain an output signal in a wider band than the input signal.

[0015]

According to the present invention, since the low-frequency component and the high-frequency component are both band-enhanced by the above-described configuration, it is possible to obtain a voice with a natural feeling and a feeling of a band, and to obtain spectral envelope information by a linear mapping function. Since the band is expanded, accurate spectral envelope information can be obtained with a small amount of calculation, and the band of the residual signal is expanded by non-linear processing, so it is possible to widen the band while maintaining the harmonic structure of the voice. Becomes Further, by performing the waveform smoothing process, temporal discontinuity of the synthesized speech can be removed, and a more natural synthesized speech can be obtained.

[0016]

An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is an overall configuration diagram of a voice band expanding apparatus according to an embodiment of the present invention. In FIG. 1, the input narrowband speech signal is analyzed by the LPC analysis unit 12 by a linear prediction analysis method, and spectral envelope information and a residual signal are extracted. The spectrum envelope information is converted into wideband spectrum envelope information by the envelope enlarging unit 13. The residual signal is converted into a wide band residual signal by the driving sound source expanding unit 14, and the LPC synthesizing unit 15 converts the output from the envelope expanding unit 13 and the output from the driving sound source expanding unit 14 into a wide band synthetic speech. LPC synthesized. The wideband synthesized speech is smoothed by the waveform smoothing unit 16 in order to reduce discontinuity of the synthesized waveform due to frame analysis. The output from the waveform smoothing unit 16 is filtered by the filter unit 17 for the frequency components missing in the narrowband input signal. The output from the filter unit 17 is added to the output from the up-sampling unit 11 that up-samples the input signal in the adder unit 18 to obtain an output signal.

A detailed description of each part will be given with reference to FIG. FIG. 2 is a detailed overall configuration diagram of the voice band expansion device of FIG. 1 (the filter unit 17 and the waveform smoothing unit 16 in FIG. 1 have a bandpass filter unit 210 and a crossfade unit as a more specific configuration. It is composed of 29).

In FIG. 2, the input narrow band speech signal is subjected to linear prediction analysis in the LPC analysis section 12 and separated into spectrum envelope information and a residual signal. In the LPC analysis unit 12, for example, the input signal is multiplied by the window function of (Equation 1).

[0020]

[Equation 1]

Here, (Equation 1) is a function indicating a Hamming window, and h (i); where i = 0, 2, ..., N-1 is a window function, and N is an analysis frame length. Then, using this h (i), the correlation function defined by (Equation 2) is calculated.

[0022]

[Equation 2]

In (Equation 2), y (i) is the input signal at time i, and r (j) is the j-th order correlation value.

Next, using the correlation value of (Equation 2), (Equation 3)
Derive the normal equation of.

[0025]

(Equation 3)

In (Equation 3), a ₀ , a ₁ , ..., A _n are linear prediction coefficients, and V _n is an nth-order linear prediction coefficient, which indicates the prediction error power when the model is constructed. . However, a ₀ = 1. Finally, the spectral envelope information can be extracted by minimizing the objective function of (Equation 4) and calculating the linear prediction coefficient from the normal equation of (Equation 3).

[0027]

[Equation 4]

The residual signal is calculated by, for example, (Equation 5) using a linear prediction coefficient that minimizes the objective function of (Equation 4).

[0029]

(Equation 5)

Ε (i) in (Equation 5) is the residual signal at time i. The spectral envelope information extracted by the LPC analysis unit 12, such as the LPC coefficient and the PARCOR coefficient, is mapped by the linear mapping function unit 23 into the same number of broadband spectral envelope information as the input using a plurality of linear mapping functions. It Further, the spectrum envelope information extracted by the LPC analysis unit 12 is referred to the narrow band codebook 21, and the vector quantization unit 22 calculates the distance by vector quantization. In the envelope addition unit 24, the wide band spectrum envelope information obtained by the linear mapping function unit 23 is converted into the vector quantization unit 2
Weighted addition is performed based on the distance calculated in 2 to obtain broadband spectral envelope information. The codebook is preliminarily trained with the narrowband spectrum envelope information, and the linear mapping function is trained with the mapping relationship from the narrowband spectrum envelope information to the wideband spectrum envelope information. To expand the band of the spectrum envelope information, for example, as shown in (Equation 6), narrow band spectrum envelope information is input, and the narrow band envelope information is linearly mapped to the wide band envelope information by the number of functions.

[0031]

(Equation 6)

In (Equation 6), z _ki is the result of mapping the narrow band envelope information x _i , where A _k is the k-th linear mapping function.

Next, from the narrow band spectrum envelope information,
By calculating the distance between each code and the narrow-band spectral envelope information using the vector quantization formula defined by (Equation 7), the weighting coefficient for each linearly mapped spectral envelope information is obtained.

[0034]

(Equation 7)

As shown in (Equation 7), the weight w _ik is calculated from the distance between the representative point V _{k of the kth} envelope information and the narrow band envelope information x _i . Using the linearly mapped spectrum envelope information obtained in (Equation 6) and the weighting coefficient obtained in (Equation 7), it is mapped to wideband spectrum envelope information by (Equation 8).

[0036]

(Equation 8)

In (Equation 8), the wide band envelope information y
_i is determined by the values calculated by the above (Equation 6) and (Equation 7).

On the other hand, the residual information extracted by the LPC analysis section 12 is up-sampled by the residual up-sampling section 25 and subjected to processing such as full-wave rectification by the non-linear processing section 26. The non-linear processing unit 26 may use non-linear operations such as full-wave rectification, half-wave rectification, square processing, and clipping. By performing these non-linear operations, it is possible to generate low-frequency components and high-frequency components of sound source information by distortion. Since the output from the non-linear processing unit 26 contains the spectrum envelope information generated by the distortion, the spectrum smoothing unit 27 smoothes the frequency characteristic by obtaining the residual subjected to the low-order LPC analysis. . For example, the order of the LPC analysis in the spectrum smoothing unit 27 is about the third order.
The output from the spectrum smoothing unit 27 has its power attenuated with respect to the input to the spectrum smoothing unit 27, and therefore is corrected by being multiplied by a constant multiple in the gain control unit 28. Alternatively, a method of performing correction by multiplying the power ratio between the input and the output to the spectrum smoothing unit 27, or L as a post-processing method
After the signals are combined by the PC combiner 15, the LPC analyzer 1
The LPC synthesis unit 15 is configured by using a bandpass filter that passes the same frequency band as the frequency band of the input signal input to
Alternatively, a method may be used in which the output is filtered, the ratio is calculated for each analysis frame so that the power of the input signal and the output of the output after the band pass filter are the same, and the ratio is multiplied.

To expand the band of the driving sound source, for example, the residual upsampling unit 25 performs upsampling of the narrowband residual signal by the calculation of (Equation 9).

[0040]

[Equation 9]

In (Equation 9), uε (i) is a signal sequence obtained by up-sampling ε (i). Next, the non-linear processing unit 26 performs full-wave rectification on the output of the residual up-sampling unit 25 using (Equation 10).

[0042]

[Equation 10]

In (Equation 10), cε (i) is the full-wave rectification of uε (i). The spectrum smoothing unit 27 performs spectrum smoothing processing by obtaining a residual signal of a low-order linear prediction method as shown in (Equation 11) from the full-wave rectified signal.

[0044]

[Equation 11]

In (Equation 11), fε (i) is a spectrum smoothed version of cε (i), and b _i is calculated using the above-described linear prediction method using cε (i) as an input signal. Is a linear prediction coefficient.

The wideband residual signal obtained as the output of the gain control section 28 and the wideband spectrum envelope information obtained by the envelope addition section 24 are LPC-synthesized by the LPC synthesis section 15 to obtain wideband synthesis. Get the sound. In the crossfade unit 29, when the shift width of the frame analysis is made shorter than the frame analysis length, this wide-band synthetic sound smoothly obtains the results obtained by overlapping in time on one time axis. Used for. A conceptual diagram of the crossfade unit 29 is shown in FIG.

In FIG. 3, a) is an input signal, and c),
e) and g) are composite output waveforms in the frame 1, the frame 2, and the frame 3, which are analysis frames that are temporally continuous, respectively. b), d), and f) are weighting functions for multiplying the composite waveform output for each frame from frame 1 to frame 3, and the output waveform h) is a waveform obtained by multiplying c) by b), and e) is d). It is calculated by, for example, adding a waveform obtained by multiplying by, and further adding a waveform obtained by multiplying g) by f). For example, one frame length is 400 points and frame shift width is 100 points.
Point, when the LPC composite output sequence at the current time is y _i and the LPC composite output sequence at the previous time is x _i , the composite output due to the crossfade is calculated as z _{i in} (Equation 12).

[0048]

(Equation 12)

The output signal from the crossfade unit 29 is output from the bandpass filter 210, for example, 0 to 300 Hz and 34.
It is filtered using an FIR filter or an IIR filter that extracts a component of 00 Hz to 7200 Hz.

The upsampling unit 11 upsamples the input signal as it is, and algebraically adds it to the output of the bandpass filter unit 210 by the addition unit 18 to obtain an output signal. In the upsampling unit 11, for example, 8 kHz
Upsample z sampling to 16 kHz sampling.

As described above, according to this embodiment, by estimating and synthesizing both the low-frequency component and the high-frequency component, it is possible to obtain a clear and natural sound with a feeling of band. Specifically, instead of directly performing pitch estimation to expand the band of the driving sound source, nonlinear processing such as full-wave rectification is performed on the upsampled residual signal to remove low-frequency components and high-frequency components. Since it is generated, it is possible to avoid the problem of generating an unnatural synthesized voice that occurs when the pitch estimation error occurs. Further, since the spectral envelope information is widened by using the linear mapping function, it is possible to generate the wideband spectral envelope information with a small number of codebooks and a small amount of calculation. Further, by smoothing the combined waveform temporally by using crossfade, discontinuity between frames due to combining can be easily flattened.

[0052]

As described above, the speech band expanding apparatus of the present invention obtains a natural synthesized sound without pitch estimation error by expanding the band without using pitch estimation in the positive (dark antonym). be able to. Further, while the amount of calculation is very large in expanding the band of the spectrum envelope information by fuzzy vector quantization, in the present invention, since the band is expanded by the linear mapping function, it can be configured with a relatively small amount of calculation and memory. Furthermore, since the present invention is a method of expanding the band of both the low frequency band and the high frequency band, it is possible to realize an excellent voice that can improve both the natural feeling by the low frequency band and the band feeling by the high frequency band. A band expansion device can be realized.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a voice band expansion device according to an embodiment of the present invention.

FIG. 2 is a diagram showing an embodiment of a voice band expanding device of the present invention.
More specific configuration diagram

FIG. 3 is a conceptual diagram for explaining an operation of a crossfade unit according to an embodiment of the present invention.

FIG. 4 is an explanatory diagram for explaining an operation of a first conventional example.

FIG. 5 is an explanatory diagram for explaining an operation of a second conventional example.

FIG. 6 is an explanatory diagram for explaining an operation of a third conventional example.

[Description of Codes] 11 Upsampling Section 12 LPC Analysis Section 13 Envelope Expansion Section 14 Driving Sound Source Expansion Section 15 LPC Analysis Section 16 Waveform Smoothing Section 17 Filter Section 18 Addition Section 21 Narrowband Codebook 22 Vector Quantization Section 23 Linear Mapping Function section 24 Envelope addition section 25 Residual upsampling section 26 Non-linear processing section 27 Spectral smoothing section 28 Gain control section 29 Crossfade section 210 Band pass filter section

Claims (6)

[Claims] 1. An LPC analysis unit for extracting spectral envelope information and a residual signal from an input signal by using a linear prediction analysis method,
An envelope expansion unit that expands the spectrum envelope information extracted by the LPC analysis unit to a desired frequency band, and the residual signal calculated by the LPC analysis unit is expanded to a desired frequency band. Using a driving sound source expansion unit for converting into a residual signal having frequency characteristics, a band expanded spectrum envelope information obtained by the envelope expansion unit, and a band expanded residual signal obtained by the driving sound source expansion unit. An LPC synthesizing unit for LPC synthesizing voice, a waveform smoothing unit for smoothing a signal sequence obtained from the LPC synthesizing unit on the time axis, and a signal obtained from the waveform smoothing unit added to the input signal A filter section for filtering the frequency band component to be converted, and a sampling frequency higher than the sampling frequency of the input signal while maintaining the frequency characteristics of the input signal. And upsampling unit for outputting a signal sequence was the voice band expansion device and an addition unit for adding the output signal from the output signal of the up-sampling unit from the filter unit. 2. The envelope enlarging unit has a linear mapping unit having a plurality of functions for linearly mapping the input spectral envelope information to a wide band spectral envelope information, a codebook having a plurality of narrow band spectral envelope information, and A vector quantizer that calculates the distance between the code in the codebook and the input spectral envelope information and performs vector quantization, and the vector quantizer at the output of each function of the linear mapping unit. The speech band expanding device according to claim 1, further comprising: an envelope expanding unit for performing weighting and adding with the output from, and converting the input spectral envelope information into wide band envelope information. 3. The driving sound source expanding unit includes a residual upsampling unit that upsamples an input residual signal,
A non-linear processing unit that performs non-linear processing on the output from the residual upsampling unit, a spectrum smoothing unit that flattens the frequency characteristics of the output signal from the non-linear processing unit, and power output from the spectrum smoothing unit. The voice band expansion device according to claim 1, further comprising a gain control unit that controls the output with reference to the output from the residual upsampling unit. 4. The waveform smoothing unit adds to each of the frame-synthesized signal sequences in order to obtain a signal sequence synthesized for each frame having temporal overlap as a continuous waveform on one time axis. The voice band expansion device according to claim 1, wherein a signal sequence smoothed on a time axis is obtained by weighting and adding. 5. The gain control unit performs gain control by filtering the observable band of the input voice from the synthetic voice and making the power of the input voice and the power of the filtered synthetic voice the same. The voice band expansion device according to claim 3. 6. An LPC analysis section for extracting spectral envelope information and a residual waveform from an input signal using a linear prediction analysis method,
A linear mapping unit that converts the spectrum envelope information extracted by the LPC analysis unit into a wide band spectrum envelope information by linearly mapping the spectrum envelope information using a plurality of functions, and a plurality of narrow band spectrum envelope information. Codebook, each code in the codebook, and the LPC
A vector quantization unit for performing vector quantization by calculating a distance from the spectrum envelope information extracted by the analysis unit, and the output of each function of the linear mapping unit, weighted by the output of the vector quantization unit. An envelope adding unit that outputs wideband spectrum envelope information by performing addition, a residual upsampling unit that upsamples the residual signal extracted by the LPC analysis unit, and an output from the residual upsampling unit. A non-linear processing section for performing non-linear processing, a spectrum smoothing section for flattening the frequency characteristic of the output from the non-linear processing section, and an output from the spectrum smoothing section with reference to the output from the residual upsampling section. A gain control unit for controlling the output from the gain control unit and the output from the envelope addition unit. An LPC synthesizer for performing LPC synthesis, a waveform smoothing unit for smoothing the signal sequence obtained from the LPC synthesizer on the time axis, and a signal obtained from the waveform smoothing unit added to the input signal. A band-pass filter section for filtering a frequency band component to be processed, an up-sampling section for outputting a signal string obtained by converting the input signal to a sampling frequency higher than the sampling frequency of the input signal while maintaining frequency characteristics, and A voice band expansion device having an output signal from a band pass filter unit and an addition unit for adding an output signal from the upsampling unit.