JPS62174798A - Voice analyzer - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Object of the Invention] (Field of Commercial Use) The present invention provides the following method based on the spectrum envelope of an input audio signal.
The present invention relates to a voice analysis device that analyzes the voice signal.

(Prior art) In speech recognition devices, vocal training devices for friends of hearing-impaired people, communication systems or speech synthesis devices using speech analysis and synthesis, input speech signals are analyzed in order to fully realize the intended processing. and its feature extraction is required. Analysis of an input audio signal is generally done based on its frequency spectrum.

This is due to the fact that human hearing is sensitive to the time-varying waveform of the audio signal itself, or rather to the spectrum of the audio signal, and recognizes signals with the same spectral shape as the same phoneme.

The voiced part of the audio signal has a structure as a periodic signal driven by vocal fold vibration. the result,
The frequency spectrum of the voiced part has a line spectral structure. On the other hand, in the silent portion, the voice signal is not accompanied by vocal fold vibration, but rather the noise caused by the airflow passing through the vocal tract is the source of the noise. As a result, this no M? The frequency spectrum of the five parts does not have a periodic structure like a line sushictor. Correspondingly, in conventional speech analysis, there are two methods: a method that assumes a periodic Norse source as the sound source of the input audio signal, and a method that assumes all noise sources. The former is an autoregressive model (AR model)
The latter is well known for its speech analysis using cepstral analysis. These audio analyzes remove fine structures from the spectrum of the input audio signal to obtain a so-called spectral envelope.

When analyzing an input audio signal using a method such as the AR model analysis method or cepstral analysis method described above to obtain the total spectral envelope, these methods assume the temporal stationarity of the system, so essentially , cannot be applied when the phoneme changes over time. Therefore, in these analysis methods, signals in a short time domain where the system does not appear to change significantly are used. A temporally quasi-stationary signal is created by cutting out the signal and applying a window function such as a Hamming window or Hanning bar to this to avoid the influence of end points. The spectral envelope obtained by completely analyzing this signal is taken as the spectral envelope at the time of signal extraction.

In addition to the above-mentioned method for obtaining a time series of spectral envelopes, a method for analyzing unsteady audio signals that assumes all temporal changes in the system as a premise of the model has also been proposed. However, in this analysis method, a time series of spectral envelopes can still be obtained as a result.

Furthermore, with the powerful analytical methods other than the above analytical methods and F7,
There is a frequency analysis method using filter puncture. This analysis method involves passing a total number of input audio signals through band filters II, each having a different center frequency, and determining the spectral quality of the filter outputs. The characteristics of this analysis method include, for example, the ease with which a hardware-based VCL can process lost time groups.

In this way, in speech analysis, the spectral envelope of an input speech signal is often determined, but the final method of analyzing the speech signal from the desired spectral envelope is to calculate its local peaks (hereinafter referred to as local peaks). Formant analysis is known in which the frequency and width of formants are all extracted from the formant. This analysis method is based on the fact that each vowel part has its own formant frequency and formant width, and in the consonant part, each consonant is characterized by the way the formant frequency changes toward the following vowel part. It is. For example, the five Japanese vowels (aiweo) are characterized by two formant frequencies Fl + F2 from the lowest frequency side, and in the voices of people of the same gender and age, FI + F2 have almost the same directness.

Therefore, by detecting the formant frequency FI+F2, it is possible to identify all vowels.

In addition, local peak analysis is also known, which extracts the spectrum envelope nolocale and focuses on its frequency and temporal transition, without focusing on full mants. This analysis method is based on the fact that phonological features are thought to appear in temporal changes in local peaks in vowel changes and consonant parts.

In addition, a method has been proposed in which the spectral envelope curve itself is used as a characteristic quantity of the audio signal and used for identification, classification, or display in subsequent processing.

As mentioned above, in the analysis of audio signals, it is important to extract the spectral envelope, and in addition to the spectral envelope itself, the frequency and width of the seven summants obtained from it, as well as the Local peak frequencies and their transitions are also used as quantities that characterize speech.

[Problems to be Solved by the Invention] By the way, when a human utters a voice, the phonology is thought to be formed by the resonance and anti-resonance characteristics of the vocal tract. For example, the resonant frequency appears as a formant on the spectrum envelope. Human beings with almost the same vocal tract structure can obtain similar spectra for the same phoneme.

However, it is known that when the vocal tract lengths are significantly different, such as between men and women, children and adults, the resonance or antiresonance frequencies shift, and the shapes of the spectral envelopes for the same phoneme do not match. There is. Therefore, in this case, the local and 7-ormant frequencies will also be shifted. This means that the purpose of analysis is to extract the same results for the same phoneme regardless of the speaker, such as speech recognition for unspecified speakers or visual display of speech for the hearing impaired. is extremely inconvenient.

Conventionally known methods for solving this problem include a method of preparing a large number of standard/mother turns, and a method of taking a ratio of formant frequencies.

In the former method, the spectral envelopes of many different friends, such as men, women, adults, and children, are registered as standard patterns, and unknown inputs and arm turns are matched to the most similar one among these many standard patterns. It attempts to recognize an unspecified number of input voices by classifying them. However, with this method, in order to correspond to any human input voice, it is necessary to prepare a very large number of standard/flat turns, and it takes a long time to compare these patterns. There are drawbacks. However, since this method does not standardize vocal tract length and extract all results, it cannot be used for the purpose of showing phoneme characteristics that are not dependent on vocal tract length.

The latter method of taking the ratio of formant frequencies is well known as a method for extracting the characteristic @ of vowels independent of vocal tract length. For further explanation of this method,
First, among the local peaks in the spectrum envelope, the first peak is considered to be relatively stable for vowels. Second. Third
extract the frequency of the formant F+ lF21F3' and the ratio between them, e.g., F+ /Fs r F2
/Fs is calculated and used as a feature quantity. If the vocal tract length increases by a times, the formant frequency increases by 17 times, that is, Fl / &, F2/ a y F 3 /
The basis of this method is that the ratio between the two is unchanged.

This method uses the first . Second. If the third fulmant is extracted stably, good results can be obtained, but if it is not stably extracted, there is a problem in that the reliability of the analysis results is significantly reduced. Well, this method is
The disadvantage is that it cannot be applied to consonant parts. In other words,
In the consonant part, the formant, which is the resonance characteristic of the vocal tract, is not defined, and in fact, the first . Second.
Since the local value corresponding to the third formant is not always observed, it is not possible to extract Fl r F2 * F3 and calculate the ratio. Please note that the full mant is not necessarily stable not only in the consonant part but also in the vowel part at the rising part and ending part, so the wrong 7 ormant frequency may be extracted. In such a case, the ratio of the fullant frequencies changes discontinuously and takes a completely incorrect value.

However, this method can only be used for stable parts of the vowel part in the speech signal, and the beginning and end of the vowel part and the consonant part must be analyzed using another method. Must-have. However, in that case, the extraction parameters are different between the stable vowel part and other parts, so it is not possible to describe the continuous change from the consonant part to the vowel part. In other words, the method of calculating the ratio of formant frequencies is essentially applicable only to stationary vowels.

As described above, a method for extracting phoneme-specific features from the spectral envelope of an unspecified number of vocal tracts with different vocal tract lengths has not yet been found.

An object of the present invention is to provide a speech analysis device that can obtain spectra specific to phonemes without being affected by differences in vocal tract length due to the gender, age, etc. of the speaker.

[Structure of the invention] (Means and effects for solving the problem) In order to achieve the above object, the present invention logarithms and normalizes the spectral envelope extracted from the input audio signal. The normalized logarithmic spectral envelope is integrated over the frequency of the logarithmic scale, and the envelope information of the spectral envelope is projected from the frequency onto the integral output.

(Embodiment) Hereinafter, a speech analysis device according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

FIG. 1 is a block diagram showing the configuration of an embodiment. Before explaining FIG. 1, the principle of the embodiment will first be explained with reference to FIGS. 2 to 6.

First, a complete comparison of the spectral envelopes of vowels according to vocal tract length is shown in Figure 2. Figure 2 shows two different vocal tract lengths (
It is the logarithmic expression of the spectral envelope P(,?'') of the same phoneme at t+)-(tt).

According to this figure 2, the frequency ranges from several 100 Hz to about 5 kHz.
In a frequency range of about
)) is the spectral envelope P2(
f) (in FIG. 2, logPg (f)) is multiplied by a constant, with the origin fixed, in the frequency axis Cf direction of the spectrum envelope P(f), which has approximately the same shape. On the other hand, from OHz to 100Hz and 5
In a frequency range of approximately kHz or higher, the difference appears significantly, and the similarity of the spectral envelopes PCf) becomes small. However, this part is related to individual differences in sound quality,
In speech analysis, it is less important. Therefore, since the vocal tract length is proportional to the resonance frequency, the vocal tract length (tt)
, (tt), the ratio (tt/ls) of spectral envelope P+(f) and spectral envelope Pg(f) is expressed by the number 100 when the magnitude is normalized and the logarithm is taken.
From Hz to about 5 kHz, the following relationship holds true. Here, the spectral envelope Pl(f)
, Pz (f) itself, but the magnitude is intended to eliminate the difference in the amplitude of the regular sign.

Now, at this time, if the first to third 7-ormants are extracted, their frequencies (Fl) - (F2), (F3)
and (F+') , (Fj') e (Fs') #i
It will look like Figure 2. In this case, between these, '1'/F1 ~ Fx'/'t -Fs'/Fs
Since there is a relationship of = r − (2), 7
The ratio (for example, as shown in the following equation (3)) of the ormant frequency CF remains unchanged.

FH/F2 SA F1'/'F'@', F
l/F3 3=p+'/p's'...
(3) This will be explained according to the embodiment shown in FIG. 3 and the conversion example shown in FIG. Figure 3 shows the temperature range of 20 to 30℃ for the five Japanese vowels.
This is an illustration of the distribution of Fl and 22 among men and women in the 20s. As is clear from Figure 3, the distributions for men and women are significantly different.For example, A for men and O for women, and A for men and O for women are distributed in the same range.

On the other hand, in FIG. 4, F1/F s + F t
/Fs distribution. According to FIG. 4, it can be seen that the difference between men and women has disappeared in the ratio of formant frequencies.

Therefore, the value on one frequency axis is multiplied by a constant within a range of approximately kHz, where the spectral envelope of the vowel is in a steady state. That is, the transformation (
R) Perform for t spectral envelope P(f).

p(f)→p'(f)=P(r-f) −(
4) At this time, the spectral envelope PCf in the invariant function space
) and P(r-, f) 't-projecting transformation (tJ) t
-If you can find it, in that space the vocal tract length <
Regardless of 1), spectral envelope PC belonging to the same phoneme
f) should have the same shape.

FIG. 5 conceptually shows this. Figure 5 shows that the difference in the spectral envelope P(f) of the Japanese phoneme “a” or “i” due to the difference in vocal tract length <1> is due to the transformation (U
) into a spectral envelope P'(f) of the same shape. That is, regarding the Japanese phoneme 17'', the spectral envelope PlACf> at the vocal tract length (tl) and the spectral envelope P2ACf) at the vocal tract length (t2) are transformed into spectra with the same shape by transformation (U). The figure shows how the envelope P'ACf) is converted into the spectral envelope P'ACf).Similarly, regarding the Japanese phonetic A, the spectral envelopes P1ICf) and P2ICf) are converted into the spectral envelope P'
This shows how it is converted to (Cf).

In this embodiment, the above conversion (U)' is realized as follows. In other words, the magnitude is normalized and the logarithmic spectrum envelope is integrated on the logarithmic scale of the frequency axis, t-Lcf
＞ Then, it becomes . Here, ε is a sufficiently small positive value close to O, and is determined by the conditions described later.

Since L(j') in equation (5) depends on the functional form of P(f), it can be written as eL,(f). If we apply the transformation of equation (4) to this, we get k==7, and logk=log h
−log rFrom this, d log k = d lo
g h−d log r However, since r is a constant, dlogk=dlogh.

...(6) becomes. If the second term on the right side of equation (6) is sufficiently small, then
LP'(f) #L, (rf>
−(7) holds true. Therefore, the spectral envelope PCf)
, Cf) is represented by the frequency Cf) in a medium-sized lameter, and considering the nine functions (p(f), t, , (f)), (
p(f), L, (f)) = (P(rf)-Lp(rf))=
(P'(f), L,'(f))
...(8), (P(f), f) → (p(f), L, (f))...
・It can be seen that the transformation (U) in (9) is a projection onto the function space that is invariant with respect to the transformation (R) in Equation (4).1, proportionally in the frequency axis direction from the vicinity of frequency O If it is expanded or contracted, the integral LCf by the logarithmic frequency axis gold hole (5)
) to absorb the deviation of the normalized logarithmic spectrum envelope on the frequency axis.

FIG. 6 is a conceptual diagram of conversion (U). Note that FIG. 6 shows the logarithmic spectral envelope loglP(j'')le instead of the envelope information binding 1 and the spectral envelope PCf), which will be described later.In this case, the above equation (8) can be expressed as the following equation Q1. be done.

(logIP(f)I,L,(f)) =(logIP'(f)l,LP'(f))...
・00 Also, if we have not the spectral envelope P(f) or the logarithmic spectrum envelope 1ogIP(f)l, but the normalized logarithmic spectral envelope loglp(f) I Kongo,
Equation (8) is expressed as the following equation α.

(logJP(f)l, Lp(f)) Here, equation (6
The conditions under which the second term on the right side of ) can be ignored are explained below. This condition is based on the fact that the practical range of the ratio (r) of the vocal tract length 1) is considered to be about 2 to 2, and the normalized logarithmic spectral envelope is approximately in the range from ε to 2ε on one frequency axis. Since it is considered to be constant, that is, approximately a constant, 0 ~ log l P (ε)・1og2 . . . is found by evaluating the integral (I)ffi expressed by the following equation (2).
・(6) The value of the audio spectrum envelope significantly decreases at frequencies less than half the pitch frequency, so if ε in equation (6) is taken to be about 1001 (Z), effectively the equation (
It can be seen from equation (6) that the second term on the right side of equation (6) can be ignored. On the other hand, if ε is set too small, the contribution from the low temperature and low frequency portions to the integral LCf shown by equation (5) becomes too large, making it sensitive to the shape of the spectrum near its origin. Therefore, the size of 5g is 1
Below 0Hz is inappropriate. From the above, the size of ε is determined as °C
is approximately 10 Hz to 100 T (approx.

Having explained the outline of this embodiment above, we will now return to FIG. 1 and explain the detailed structure for carrying out the processing of this embodiment.

In FIG. 1, the spectral envelope extraction unit 11 extracts the spectral envelope PCf)'l: of the input audio signal (A!N). Here, the method for extracting the spectral envelope is as follows:
A variety of methods can be used, such as a spectral envelope extraction method in a speech analysis method using the human R model, a spectral envelope extraction method in a speech analysis method using Kegnutrum analysis, and a spectral envelope extraction method in a speech analysis method using frequency analysis using filter puncture. is possible.

The logarithmization unit 12 logarithms the magnitude of the spectral envelope P(,I”) extracted by the spectral envelope extraction unit 11. The normalization unit 13 logarithms the magnitude of the spectral envelope 1oglP(f ) normalize the size of l.

The logarithmization section 12 and the normalization section 13 together form a conversion section 10f. Here, the logarithmic spectral envelope 1oglP(f)I
or, for example, by automatic gain control, the logarithmic spectral envelope 1oglP(f)
Examples include a method of differentiating l with respect to frequency (f), dropping the °C constant term, integrating it, and adding a constant °C value.

The integrating unit 14 receives the normalized logarithmic spectrum envelope 1ogIP(f)I? output from the normalizing unit 13. , the frequency on the logarithmic scale is used as a variable, and the ′'' CC-integration is performed.In other words, this integration section 14 converts the spectrum envelope 10g1P(f)1 into the above equation (
5), the temperature is integrated according to the integral function shown in 5). Here, "'C30Hz" is used as the value of ε.

The projection unit 15 receives the logarithmic spectrum envelope logIP(f)l output from the logarithmization unit 12 and the integration result output from the integration unit 14, and calculates the logarithmic spectrum envelope 1ogI
P(f)I k As shown in Figure 6, project from the frequency (f) onto the integral function L(f) (=t, , (f)),
The projection results are shown below. In other words, this projection section 15 is
The input audio signal (AlN) is expressed on the X-axis of the X-7 orthogonal coordinate by taking (7)' and the logarithmic spectrum envelope LogIP(f)lk on the y-axis, and displaying it in a linear meter according to the frequency Cf). The analysis results are converted into gold patterns.

In addition, in the processing of the projection unit 15, the value taken on the y-axis is used, or as is clear from the previous explanation of 2αQ and αη, the spectral envelope P(f) and the normalized logarithmic spectral envelope 1ogIP(f )l may be used. Mayu, not limited to this, the normalized spectral envelope P(f) P
(f) may be used. In other words, in this invention, the envelope information provided for projection includes at least the four points mentioned above. means turn.

In addition, in the processing of the projection unit 15, it goes without saying that the envelope information may be taken on the X axis and Cf)'jr on the y axis.

Here, an example of actual measurement using voice analysis in this embodiment will be explained.

Figures 7 and 8 show the logarithmic spectrum envelope tag 1p(f) I of the female and male Japanese phoneme “i”, respectively.
'frshow. This is the logarithmic spectral envelope.

First, the audio signal (A□N) input to the extraction unit 11 is
The audio sampled with a condenser microphone was sampled at a cycle of 50μ degrees and digitized using 't12 pit. In addition, to collect this audio, I used Word's unique memory at 8°C.

The extraction unit 11 calculates the spectral envelope PCf) from this input audio signal (A, N) by cepstral analysis.
There is. The kegstral analysis is set to obtain the spectral envelope P<f) by differentiating the frame of 1024/India length of the vowel stable part, applying a Hamming window to this, and Fourier transforming the resultant using the FFT algorithm. There is.

The logarithmization unit 12 takes the logarithm of the absolute value of the spectrum envelope PCf) obtained as described above, performs an inverse 7-lier transform on it to obtain a cepstrum, and then performs a cutoff on the quefrency by 1 By applying a rectangular window of .7 to 2.5 m WMt and performing full Fourier transformation, the logarithmic spectral shape log 1p(f)l is obtained.

Note that this logarithmic spectrum envelope log I P (f)
1, the cutoff on the quefrency is selected in accordance with the pitch frequency. Also, in order to normalize the magnitude of the logarithmic spectral envelope log IP(f) l, this spectral envelope log IP(f) It
! , it is calculated after fixing the O-th missing component of the kegstrum to a constant value.

Logarithmic spectral envelope log shown in Figures 7 and 8
IP(f)I is obtained as described above.

Comparing the logarithmic spectral envelopes in Figures 7 and 8, the spectral shapes of the two are similar in the frequency range below about 5 kHz, but in the frequency range above this, the female's spectral shape is more similar. is shifted higher.

For this logarithmic spectral envelope log IP(f)l, it is expressed as g t-50Hz by equation (5), or (
f), set it on the X axis, and calculate the logarithmic spectrum envelope l
og IP(f)l ky axis is shown in FIGS. 9 and 10, respectively. Looking at this, the peak height,
Although the details are slightly different, Figures 7 and 8
It can be seen that the deviation in the frequency direction at 5A has been eliminated.

FIG. 11 shows the time-series changes in the logarithmic spectrum envelope log IP(f)I obtained by changing the frame position in degrees Celsius and temporally changing the frame position for the rising part of the phoneme "Yu" uttered by a woman. FIG. 12 shows how to apply the transformation (U) of this embodiment to the logarithmic spectrum envelope log IP(f)l of FIG.

According to the measurement results, it can be seen that the conversion (U) of the present example has a positive effect on the consonant part and the rising part of the vowel, and the spectrum changes are displayed smoothly.

Figures 13 and 14 respectively show the time-series changes in the Japanese phoneme "a" for a man and a woman obtained in the same manner as in Figure 12. Looking at the vowel parts in Figures 13 and 14 and Figure 12 above, it can be seen that the large difference in vocal tract length between men and women has been eliminated.

As described in detail above, this embodiment uses envelope information (p(f
), p(f)-log IP(f)l or log
The voice analysis result is obtained by projecting IP (f) I) onto (f) using the transformation (U) defined by equations (5) and (9). It is composed of

According to such a configuration, it is possible to eliminate differences in analysis results due to differences in the vocal tract lengths CL) of speakers, and it is possible to always obtain phoneme-specific analysis results. In this case, this embodiment can be applied to the spectral envelope of any part of the input audio signal (AIN), regardless of whether it is voiced or unvoiced or whether it is a vowel or a consonant. Furthermore, since the analysis result does not depend on the extraction precision of the formant frequency (F) or its stability, it can be applied to all parts of the input audio signal (A, N).

In particular, according to this embodiment, the process of transitioning from the consonant to the vowel can be realized without being affected by differences in individual vocal tract lengths (t), which was previously impossible. The problem of determining changes in spectral envelope can be solved.

In addition, in this embodiment, the spectral envelope P(f) and the logarithmic spectral envelope l are used as the integrand of equation (5).
og IP(f) Instead of I, the normalized logarithm <- can remove the influence of loudness on phonology.

Note that the present invention is not limited to the above embodiments.

For example, in the previous embodiment, in order to obtain the integrand of equation (5), the vector envelope PCf output from the extraction unit 11 is
)'fr We have explained the case of logarithmization and normalization after t, but
Of course, the reverse process may also be performed.

Furthermore, it goes without saying that the speech analysis device of the present invention can be realized by either hardware or software.

Next, another embodiment of the speech analysis device according to the present invention will be described in the fifteenth embodiment.
A detailed explanation will be given with reference to all of FIGS. 18 to 18.

First, the configuration of the embodiment will be explained with full reference to FIG. 15.

This embodiment, like the previous embodiment, includes a spectral envelope extraction section 11', a logarithmization section 12', a normalization section 13', an integration section 14', and a synthesis section 15'. Here, the combining section 15' is the projection section 15 in the previous embodiment.
Perform all the same operations as . The speech analysis device in this embodiment further includes a control device 16' and a display device 17'.

Transformation 910' is performed by the logarithmization unit 12' and the normalization unit 13'.
f form.

Next, the construction of each part will be explained in detail with reference to FIGS. 16 and 17.

The spectral envelope extraction unit 11' includes an adjuster 211 for adjusting the level of the input audio signal, a high frequency emphasizing circuit 212 for emphasizing the high frequency range of the input audio signal, and a circuit 212. A buffer 213 for reinforcing all output signals for subsequent stages; a distribution circuit 214 for dividing the signal output from the buffer 213 into appropriate frequency ranges, and reinforcing and outputting each; filter puncture 21 for extracting each predetermined frequency range according to the output and creating a spectrum envelope;
5, and a multi-blefter 216 for serially outputting the spectral envelope data output from the filter puncture 215 in accordance with detailed control data for output selection from the control device 16'.

In this embodiment, the distributor 214 includes four distributors 214-1
It consists of 214-4. Each of the filter punctures 215 consists of eight filter sections 215-1 to 215-II, and each filter section, as shown in FIG.
It is composed of a band pass filter 315, a rectifier 316, and a low pass filter 317. Here, the bandpass filter 315 and the low/4th filter 317 are each DT-2 manufactured by NF Circuit Design Block Co., Ltd.
12D and DT-6FL1 are used.

The bandpass filter 315 is controlled according to control data from the control device 16' and passes signals of a specific frequency band. The rectifier 316 includes a bandpass filter 315
The signal output from the filter is full-wave rectified and supplied to a low-chair filter 317. Low/4', X filter 317
extracts the low frequency component from the signal from the rectifier 316, that is, extracts the waviness component. Thereby, a spectral envelope in the fractional band limited by the pantone filter 315 is obtained.

The logarithmization unit 12' sequentially inputs all the spectrum envelopes converted into serial data output from the spectrum envelope extraction unit 11, samples the spectrum envelope data of each channel, and holds the spectrum data in degrees Celsius. sampling/holding circuit C8/H) 222 and S/H22
2, and a logarithmization circuit (LOG) 223 for inputting and logarithmizing a signal held at 2.

Logarithmization part 1! is followed by a normalization section 13', and a logarithmization section 1!
The logarithm (cutram envelope) output from the normalization circuit 23
) is normalized. The output from the normalization circuit 23 is supplied to the integrator 241 of the integrating section 14' and the combining section 15'.

The synthesizing section 15' includes a sampling/holding circuit (S/H) 251 for sampling and holding the data integrated by the integrating section 14';
The A/D converter 252 for converting the data held and processed by 1 into digital data and the clock pulse from the pulse generator 257:
A buffer memory l) 253 for temporarily storing data output by the D converter 252, and a normalization unit 1
Sample the spectrum envelope output from 3,
A sampling/holding circuit (S/H) 254 for holding, and an A/H circuit for converting the data held by the S/H 251 into all digital data.
A buffer memory 256 for temporarily storing data output by the A/D converter 252 and a buffer memory 256 for temporarily storing data output from the A/D converter 252 according to clock pulses from the D converter 255 and the nogle generator 257 Although the input audio signal can be detected with the above configuration, in which the data stored in the memory 253 is stored as an address, in order to display all the detection results, in this embodiment, the display section 17' is further provided with a synthesis section. 15'.

The display section 17' has a ninth display controller 259 for reading all the data stored therein from the system memory m, and a display controller 259 for reading out all the data stored therein from the system memory 4 and storing the data for display. frame memory 258;
and a CRT 2 for displaying data ft in the frame memory 258 according to instructions from the controller 259.
60.

Here, the frame memory 258 has eight frames.
/ Consists of one memory. The friend data read from the system memory 4 is sorted into eight stages, and stored in each frame memory according to the sorting. This causes the data to correspond to one point in some frame memory,
It becomes possible to display the obtained spectral envelope in gradations.

Next, the operation of the audio signal analysis device having the above configuration will be explained.

Before the audio signal is input, the BCD is sent from the control device 16'.
Control data for the code is supplied to a pantone filter 315 and a low pass filter 317.

In this embodiment, there are 64 band-pass filters Fi, and when 64 of the frequency bands of the spectrum envelope match the logarithmic frequency axis, the finally obtained spectrum envelope is used for comparison on the frequency axis. be able to. Remember, when the control data matches the Mel scale axis, it can be compared on the Norr scale, which corresponds to human hearing.

76' Control device 1; After the bandpass filter 315 and the low-pass filter 317 are set according to the control data from 7', the audio signal is input to the regulator 21). The input audio signal is adjusted to have an appropriate DC level. The adjusted signal is input to the high frequency emphasis circuit 212. Since the audio signal attenuates as it reaches higher frequencies, the circuit 212 emphasizes the higher frequency components so that they are at approximately the same level as the lower frequencies, and outputs them to the buffer 213.

In this example, the spectrum envelope to be analyzed is 64
Since it is divided into the frequency range of the channel and detected,
The audio signal must pass through a number of filters and is augmented by a buffer 213. Furthermore, buffer 2
The friend signal output from 13 is supplied to a distributor 214,
It is reinforced and divided into four bands. Distributor 214-
The signal output from 1 is the filter puncture FBOA, ?
15-1 and FBOB 215-2. The same applies to others.

Each filter puncture 215 includes eight channels;
Each channel consists of a bandpass filter 315, a rectifier 316, and a low frequency filter 317, as shown in FIG. The filter 315-1 is supplied with BCD restriction 1 data from the outside, and according to the data, the total rectification circuit 316-1 converts the predetermined frequency components of the input signal at °C.
Output to. The signal full-wave rectified by the rectifier circuit 316-1 is supplied to a low-pass filter 317-1.

The low-pass filter 317-1 also has its band controlled by external BCD data, and extracts a predetermined frequency component from the ratio signal output by the rectifier 316-1. As a result, a spectral envelope in the band limited by the bandpass filter 315 is obtained.

The outputs of low-pass filters 317-1 to 317-16 are supplied to multiplexer 221-1. Similarly,
The outputs from filters 317-17 to 317-32 are sent to multiplexer 221-2, the outputs from filters 317-17 to 317-32 are sent to multiplexer 221-2, and the outputs from filters 317-17 to 317-32 are sent to multiplexer 221-2. It is supplied to multiplexer 221-2.

In this example, 64 channel bands are selected, so 64 spectral envelopes are obtained.
Multiplexers 221-1 to 76' 221-4 sequentially output data for each channel based on timing control signals provided from the controller pieces.

The signal output from the multiplexer 221 is logarithmized, but at this time, the signal output from the multiplexer 221-1 contains noise at the beginning of the signal, and after it stabilizes, it is logarithmized by the S/H 222. A timing control signal is output to be sampled and held. The signal held by S/H222 is LO
Logarithmized by G 223.

The logarithmized nine-sector envelope signal is output to an integrator 241 and a combining section 15'. The integrator 24 integrates the input spectrum envelope f:
Output to 1.

In the synthesis section 15', the S/Hs 251 and 254 sample and hold the input spectrum envelope in synchronization with the timing of the output from the multiplexer 22.
Output to A/D Hun/Guichiyo 252 and 255, respectively. A/D converters 252 and 255 convert the input analog signals into digital data and output the digital data to buffer memories 253 and 256.

Memories 253 and 256 are supplied with a clock from a pulse generator 257. The generator 257 76' is supplied with a signal h-C synchronized with the timing signal supplied to the multiplexer 216 from the control device N', and based on this, the generator 257 stores the clock as address data in the memory. 253 and 256. Therefore, the integral property is temporarily stored in the memory 253, and the logarithmized value of the vector envelope is temporarily stored in the memory 256 at the same address as the memory 2253.

The data temporarily stored in the buffer memory 256 is stored in the controller 259 in the system memory. At this time, the buffer memory 253
The data stored in the buffer memory 256 is the integrated data. Data is not necessarily stored. In the extraction process, the address will be skipped first.

/B/ When the address is different, it is also possible to supply interpolated data from the control device β', read it out by the controller 259, and supply it to the frame memory 258. At this time, it is determined which of the eight levels the read data corresponds to, and the data is supplied to one frame memory 258 based on the result. The data stored in the frame memory 258 is transferred to the CRT 26 o
will be wisely shown. At this time, the data is divided into eight levels, so it is shown as 1st floor AQ a. For example, the parts corresponding to the peaks of the time evolution of the spectrum envelope are displayed in white, and the parts corresponding to the valleys are displayed in black. As described above in detail, in this embodiment, if all the control data from the control device 7 and t' are appropriately selected and supplied to the filter bank. Integration can be done on the standard axis.

[Effects of the invention] According to this invention, it is possible to obtain a phoneme-specific analysis pattern without being affected by the length of the vocal tract. However, if the same analytical results are obtained, it can make a significant contribution.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing the configuration of one embodiment, Fig. 2 is a spectrum envelope diagram for explaining the entire outline of one embodiment, and Fig. 3 is a block diagram showing the configuration of one embodiment.
FIG. 4 is a measurement diagram for completely explaining the outline of one embodiment,
5 and 6 are regular spectral envelope diagrams for explaining the outline of one embodiment, and FIGS. 7 to 14 are companion spectral envelope diagrams for explaining actual measurement examples of one embodiment. FIG. 15 is a block diagram showing the configuration of another embodiment, FIGS. 16 and 17 are block diagrams fully explaining the details of each block shown in FIG. 15, and FIG. 18 is a block diagram showing the details of each block shown in FIG. 16. Show details of filter puncture. 10... Conversion section, 11... Spectrum envelope extraction section. 12... Logarithmization section, 13... Normalization section, 14...
Integral part. 15... Projection section, 10'... Conversion section, 11'...
Suictor envelope extraction section, 12'... Logarithmization section, 13'
... Normalization section, 14'... Integration section, 15'... Projection section, 16'... Control device, 17'... Display section. Applicant's representative Patent attorney Takehiko Suzue Figure 1 Figure 2 Lp(+) Figure 9 zoglp(t)l Male: V Figure 10 Figure 1z Figure 18

Claims (15)

[Claims] (1) A conversion means for converting the input spectral envelope so that its value becomes appropriate, and inputting the spectral envelope converted by the conversion means, and integrating the input spectral envelope with respect to a predetermined variable. and a projection means for inputting the spectral envelope transformed by the converting means and the spectral envelope integrated by the integrating means and projecting the spectral envelope regarding the integrated spectrum. A voice analysis device featuring: (2) The conversion means includes a logarithmization means for logarithmizing and outputting input spectral envelope data, and a normalization means for normalizing the spectral envelope logarithmized by the logarithmization means. A speech analysis device according to claim 1, characterized by comprising: (3) The conversion means includes normalization means for normalizing the input spectral envelope, and logarithmization means for logarithmizing the spectral envelope normalized by the normalization means. A speech analysis device according to claim 1, characterized in that: (4) The speech analysis device according to claim 1, wherein the conversion means includes a logarithmization means for logarithmizing the input spectrum envelope. (5) The speech analysis device according to claim 1, wherein the conversion means includes normalization means for normalizing the input spectrum envelope. (6) The speech analysis device according to claim 1, wherein the integration by the integrating means is performed with respect to frequency as a predetermined variable. (7) Spectral envelope extraction means for extracting a spectral envelope from an input audio signal according to an input first control signal, and outputting data representing the extracted spectral envelope according to an input second control signal. and converting means for converting the spectral envelope extracted by the spectral envelope extracting means so that its value becomes appropriate; and inputting the spectral envelope converted by the converting means, an integrating means for integrating an input spectrum envelope; and inputting the spectrum envelope converted by the converting means and the spectrum envelope integrated by the integrating means according to an input third control signal, and calculating a spectrum related to the integrated spectrum. outputting the first and second control signals to a projection means for projecting an envelope and the spectral envelope extraction means;
A speech analysis device comprising: control means for outputting the third control signal to the projection means. (8) The conversion means includes a logarithmization means for logarithmizing and outputting input spectral envelope data, and a normalization means for normalizing the spectral envelope logarithmized by the logarithmization means. A speech analysis device according to claim 7, characterized by comprising: (9) The conversion means includes normalization means for normalizing the input spectral envelope, and logarithmization means for logarithmizing the spectral envelope normalized by the normalization means. A speech analysis device according to claim 7, characterized in that: (10) The speech analysis device according to claim 7, wherein the conversion means includes a logarithmization means for logarithmizing the input spectrum envelope. (11) The speech analysis device according to claim 7, wherein the conversion means includes normalization means for normalizing the input spectrum envelope. (12) The speech analysis device according to claim 7, wherein the integration by the integrating means is performed with respect to frequency as a predetermined variable. (13) The speech analysis device according to claim 7, wherein the integration by the integrating means is performed with respect to a Mel scale as a predetermined variable. (14) The spectral envelope extraction means includes a high-frequency emphasizing means for emphasizing the high-frequency range so that the high-frequency level of the input audio signal approaches the low-frequency level; band-pass filter means for dividing into a predetermined plurality of channels and performing band-limiting according to the first control signal; and full-wave rectification of the audio signal of each channel band-limited by the band-pass filter means. and according to the first control signal,
low-pass filter means for detecting a band-limited spectral envelope from the audio signal front-wave rectified by the rectifying means; 8. The speech analysis device according to claim 7, further comprising multiplexer means for selectively outputting the output according to the control signal. (15) The projection means includes a first A/D conversion means for converting data from the integration means into digital data, and the data converted by the first A/D conversion means into the first A/D conversion means. a first buffer memory means for temporary storage and a second A/D for converting the data from the converting means into digital data according to the control data of No. 3;
a conversion means, a second buffer memory means for temporarily storing the data converted by the second A/D conversion means in accordance with the third control signal (2), and the second buffer memory. Claim 7, further comprising a storage means for storing the data stored in the means as an address of the data stored in the first buffer memory means. Speech analysis device.