JPH03236100A - Voice detection system - Google Patents

Abstract

PURPOSE:To reduce an omission and the addition of a noise by calculating feature parameters from a voice signal and projecting in a specific main component vector space, discriminating whether a voice signal is a voiceless or voiced sound according to the position of the projective point, and increasing voice/ voiceless-sound detection accuracy. CONSTITUTION:The voice signal is inputted and encoded by a voice encoder 41 and a noise encoder 42. One of the outputs of the encoders 41 and 42 is inputted to a cell generating circuit 46 through a selector 45 and made into a cell. The selector 45 is switched according to the discrimination result of a voiceless/voiced-sound discriminating device 43 and only the voiced section of the voice signal is made into the cell. A noise detector 44 detects only the voiceless sound section, wherein a background noise is encoded by the noise encoder 42. The selector 45 is switched to the side of the noise encoder 42 when the discriminating device 43 detects the voiceless sound. The device 43 calculates the LPC cepstrum C of the input voice signal to analyze the main component, an inner product calculator 51 finds the projection point Q of C, and the discriminating device 56 decides whether a frame is voiced or voiceless according to the Q.

Description

[Detailed description of the invention] [Object of the invention] (Industrial application field)
The presence/absence of voice signals, which is the basic technology of TM communication and voice recognition,
Regarding silence detection method.

(Prior Art) If a device that processes only the voiced portion of an audio signal does not accurately distinguish between silence and voice, the transmitted voice will be interrupted and the error rate of voice recognition will increase. Also, A
In the TM communication system, lines cannot be used effectively. Therefore, very accurate silent/sound discrimination is desired. Therefore, a silence/voice identification device that does not depend on signal level fluctuations due to changes in the usage environment and can reduce detection omissions even when the ambient noise level is high or for low-level voice signals such as word order consonants. A conventional example of this is JP-A-Sho ft (1
-2QQ3G (described in No. 1).

FIG. 1 is a block diagram of this conventional device. An audio signal input via a microphone or the like is supplied to an energy extraction circuit 5 and a spectrum extraction circuit 6.

The energy extraction circuit 5 is composed of a rectifying and smoothing circuit, and extracts the power (logarithmic power) as a characteristic parameter of the audio signal every frame period of p predetermined time in units of silent/sound discrimination time. The spectrum extraction circuit 6 has a low frequency range (250-6
00Hrl, middle range (6110-1500Hrl high range (15
It consists of three types of bandpass filters of 00 to 4000 tl and three rectifying and smoothing circuits connected to their output terminals, and also calculates the power (logarithmic power) of each band as a characteristic parameter of the audio signal for each frame period. Extract.

The energy extraction circuit 5 and the frequency of spectrum extraction 6 constitute a feature amount extraction section 13.

The outputs of the energy extraction circuit 5 and the spectrum extraction circuit 6 are supplied to a multiplexer 7, and the signal power from the energy extraction circuit 5 and the power for each band from the spectrum extraction circuit 6 are time-divisionally divided into a silence/speech discriminator 8. is supplied to Identification 8 determines whether each frame of the audio signal is silent or has a sound (a sound frame includes a voiced sound frame and an unvoiced sound frame). A threshold memory 9 and a standard pattern memory 10 are connected to the discriminator 8 . The threshold value memory 9 stores two threshold values E1 and E2 for determining whether there is no sound or sound based on the power. The standard pattern memory 10 contains coefficients of a linear discriminant function for detecting whether a frame to be detected is a silent frame or a voiced frame, and a coefficient for detecting that a frame to be detected is a silent frame or a voiced frame. Stores the coefficients of the linear discriminant function. These threshold values and coefficients are determined and stored in advance based on the statistical properties of audio signals including silence, voiced sounds, and unvoiced sounds generated under the environment in which this silence/speech identification device is used.

The identification result of the classifier 8 is input to the sound period start and end candidate detector 11, and based on the identification result for each frame,
Detect starting and ending candidates (frames) of a sound period. The detection result is sent to the sound period detection unit 12, where the start and end of the sound period are finally determined.

The operation of this conventional example will be explained. The audio signal of each frame is converted into power LPW by the energy extraction circuit 5, and the power LPi for each band is converted by the spectrum extraction circuit 6 (i is a parameter indicating the band, here, i=1 to 3).
is converted to The silence/speech discriminator 8 uses these four feature parameters LPWSLPi, thresholds E1 and E2, and coefficients of the linear discrimination function stored in the memory 9.10 to determine whether the frame is voiced or not. Determine whether

This determination is first made by comparing two threshold values El and E2 with the power LPW as follows.

If LPW>El, there is sound, L PW>E
If 2, there is no sound, E2≦LPW≦E1
If so, it is determined to be indeterminate.

If it is determined to be undefined, further determination is made based on the discriminant function value FX expressed as follows using the power LPi for each band and the coefficients of the two linear discriminant functions.

1=1 Here, Ai is the coefficient of the discrimination function stored in the memory IO, LPi is the standard pattern, and the standard pattern is also obtained by statistical processing of the audio signal uttered in the usage environment in advance. .

The function value FX is determined to be negative when there is no sound, and positive when there is a sound consisting of an unvoiced sound or a voiced sound.

FX when a coefficient of a linear discriminant function is used to detect silence or voiceless sound as the coefficient At, and FX when a coefficient of a linear discriminant function is used to detect silence or voiced sound. are calculated respectively, and if either one takes a positive value, it is determined that there is a sound, and otherwise it is determined that there is no sound.

In this conventional example, silence/voice discrimination is performed based on the difference between the spectral shape extracted by the spectrum extraction circuit 6 and the standard spectral shape of silence, unvoiced sounds, and voiced sounds. It is possible to prevent omission of identification of voiced consonants, voiced consonants, etc. However, the characteristic parameters related to the spectral shape are low frequency (25G
-600H, midrange (6GG-15flOHI),
Power for each of the three high-frequency bands (1500-4000) is used, but there is no theoretical basis for selecting this feature parameter, and the number of feature parameters is small, so it is difficult to distinguish between silence and sound. If a mistake is made, failure to detect a voice signal or addition of noise may be unavoidable. For example, the spectrum of unvoiced sound and the spectrum of noise are each
In the case shown by the solid line and broken line in the figure, Ai (i=1 to 3) even though the shapes of the two are greatly different.
When =1, the function value FX of equation (1) will be the same for unvoiced speech and noise, resulting in an erroneous determination. This is because the number of characteristic parameters that define the spectral shape is small and the selection of the parameters is not appropriate. Furthermore, since there is no theoretical basis for selection, selection must be made by trial and error, and despite a great deal of effort, the resulting feature parameters are not necessarily appropriate. be.

Incidentally, if the number of parameters is increased, erroneous judgments can be reduced, but the amount of calculation to obtain the discriminant function value of equation (1) increases. Furthermore, although the above-mentioned publication states that the Mahalanobis distance can be used instead of the discriminant function in equation (1), there is a problem in that the amount of calculation increases even more if it is used.

(Problems to be Solved by the Invention) As mentioned above, in the conventional voice/silence detection method, when the number of parameters is reduced in order to reduce the amount of calculation, the voice/silence determination is incorrect and the voice is dropped. There is a problem that the addition of noise and noise may be unavoidable. Further, the conventional method has the problem that it requires a lot of effort because there is no theoretical selection criterion for selecting parameters.

The present invention has been made in view of these problems, and an object of the present invention is to provide a speech presence/non-speech detection method that has high precision in detecting speech presence/non-speech and is less likely to drop out voices or add noise.

[Structure of the Invention] (Means for Solving the Problems) The present invention labels sounds collected in advance in an environment where a telephone or a recognition device is used as sound or non-sound by preview listening or visual confirmation of waveforms. Next, the characteristic parameters of the sound part and the silent part are obtained, and principal component analysis is performed on each of them to obtain the principal component vectors of the sound part and the silent part in advance.

Next, the feature parameters of the detection target frame are projected onto the principal component vector space of the sound part feature parameters or the principal component vector space of the silent part feature parameters, and the position of the projection point and the temporal change amount of the projection point or The presence or absence of sound is determined based on the amount of change in the projection point over time.

(Function) When using the voice detection method of the present invention as described above, first, characteristic parameters of an acoustic signal such as a voice signal are determined. Next, consider reducing the number of parameters after converting that parameter into another parameter than the original feature parameter. Figure 3 shows this concept. In Fig. 3, L original feature parameters are expressed as Xi (i=l, 2...
, L) and let X be a vector whose element is Xi. The transformation is orthogonal transformation, and the transformation matrix is A. The feature parameters after conversion are yi (i=1.2...,L), y
A vector with i as an element is defined as YSN parameters yi(
j=1.2. ..., N) is the feature parameter vector Y' with the remaining (L-N) pieces being zero (however, N<L
). At this time, the error vector e caused by the reduction in the number of parameters is described as the difference between the original feature parameter vector X and the inverse transformation of Y' as shown in the following equation.

e=X-A-"Y- = A-" (Y-Y-) The root mean square value of this error σr' = E [e' The number of feature parameters can be reduced by performing the transformation that minimizes el. The error due to this is minimized. However, it is known that the variable that minimizes 7r2, where E is the expected value, is a transformation using a matrix A whose row vectors are the eigenvectors of the autocorrelation matrix of Xi, that is, KL transformation. Further, the eigenvectors are the same as the principal component vectors obtained by Xi principal component analysis, and the eigenvectors corresponding to the larger eigenvalues correspond to the first, second, third, . . . principal component vectors.

After performing KL transformation on the L feature parameters X, the operation of reducing the number of parameters corresponds to projecting X onto an N-dimensional principal component vector space whose coordinate axes are the first to Nth principal component vectors. Therefore, by projecting the feature parameters onto the principal component vector space, we can reduce the number of feature parameters while minimizing the error in expressing the original feature parameters with fewer parameter dimensions, in other words, the loss of information in the original feature parameters. can be reduced.

The feature parameters of the sound part and the silent part are distributed in a specific region on the principal component vector space due to differences in characteristics, for example, differences in spectral shapes. Utilizing this property, the presence/absence of speech is determined by comparing the projection point when the feature parameters are projected onto the principal component vector space with a preset region of speech/non-speech.

In addition, the projection point on the principal component vector space of the feature parameter shows a certain characteristic change over time corresponding to the temporal change in the nature of the voice, such as from silence to voice, and from voice to silence. do. Utilizing this property, the presence/absence of sound is determined by comparing the parameter representing the temporal change of the projection point with a preset threshold, or by matching the pattern of the parameter representing the temporal change with a preset standard pattern. be able to. The voice detection is performed by the sound/non-sound determination based on the temporal change of the projection point alone or in combination with the above-described voice/non-sound determination based on the position of the projection point.

This method provides high accuracy in detecting speech/non-speech, and reduces omission of speech and addition of noise.

(Embodiment) An embodiment of the silence/speech identification device according to the present invention will be described below with reference to the drawings. FIG. 3 shows an outline of the invention.

First, feature parameters are obtained from the audio signal using a known method. In addition to the conventional spectrum, characteristic parameters include LPC cepstrum, signal power, number of zero crossings, linear prediction coefficient, autocorrelation function, DFT coefficient, etc.
There are also combinations of these. Here, it is not necessary to select the number, types, etc. of feature parameters, and it is desirable to obtain as many parameters as possible. Next, consider reducing the number of feature parameters so as not to affect the silence/speech discrimination accuracy. To this end, in this invention, the number of parameters is reduced after first converting the feature parameters into other parameters, and this conversion involves inversely transforming the reduced number of transformed parameters and the original parameters. Conversion is performed so that the error with the parameters is minimized.

Specifically, as shown in FIG. 3, it is assumed that L original feature parameters are Xi, (i=1 to L), and a vector having Xi as an element is X. The transformation is orthogonal transformation, and the transformation matrix is A. The feature parameters after conversion are yi (i
-1 to L), where Y is a vector whose elements are yi, and N parameters yj of the feature parameters yi after conversion are
(j = 1 to N) and the remaining (L-N) parameters are set to 0.
shall be. However, it shall be NIL. At this time, the error vector e caused by reducing the number of parameters is calculated as the difference between the original feature parameter vector expressed.

e=X-A-'Y" = A-1(Y-Yi...... (2) Perform the transformation that minimizes the root mean square value of this error (rr2=E [e' e] If we do this, the error due to reducing the number of feature parameters will be minimized, where t is the transpose of the matrix E[
] is the expected value.

This is known as the KL transformation, which is a transformation using a matrix A whose row vectors are the eigenvectors of the autocorrelation matrix of the transformation parameter Xi that minimizes the root mean square value expressed by equation (2). The eigenvectors are the same as the principal component vectors obtained by principal component analysis of Xi, and the eigenvectors corresponding to the larger eigenvalues correspond to the first, second, . . . principal component vectors.

Letting the feature parameter vector be Xi (i=1 to M), the root mean square value of equation (2) for each vector Xi!
The transformation that minimizes j' r ' is defined by the eigenvector obtained by principal component analysis of the autocorrelation matrix of the feature parameter vector expressed by the following equation.

Here, xi, , x, . . . xit are elements of the feature parameter vector xi.

From equation (3), it can be seen that the autocorrelation matrix R is the average of the autocorrelation matrix Ri of each feature parameter vector Xi shown in the following equation in an L2-dimensional space, that is, the center of gravity.

・・・・・・・・・(3) ・・・・・・・・・(4) Autocorrelation matrix Ri of M feature parameter vectors Xi
(i=1.2,...M) is represented by one autocorrelation matrix R, the autocorrelation matrix R minimizes the root mean square error E between Ri and R defined by the following equation There will be a queue.

LM L L E= Σ Σ Σ (R(k, 1
) −RI (k, J) ) M 1・l k=I
I=1 Because, by partially differentiating the above equation with T(k,/) and setting it as O, the following equation is obtained, which is
This is because it becomes R. However, R(k, l)Ri (
k, l) are the autocorrelation matrices R and Ri (k,
47) It is an element.

Principal component analysis can be performed by performing KL transformation on the feature parameter vector Xi using such a transformation matrix R.

Here, L feature parameters xi (i=1 to L)
After performing KL transformation on , the operation of reducing the number of parameters corresponds to projecting the feature parameter vector X onto an N-dimensional principal component vector space having the first to Nth principal component vectors as coordinate axes. Then, by projecting onto the principal component vector space, the number of original feature parameters can be reduced while minimizing the error caused by reducing the number.

FIG. 4 shows a note ossifier 42 which is an embodiment of the silent/sound discriminating device according to the present invention, which is constructed based on such a principle.
and encoded there.

The encoding speeds of the encoders 41, 42 are different, with the speech encoder 41 being faster. The encoding system of the audio encoders 41 and 42 is ADPCM (AdspMCDi).
Eyerelixl PIIgICode Mod++1
xtion) is supplied to the encoding channel 46, and the audio signal is converted into cells. Switching of the selector 45 is performed based on the discrimination result of the silent/speech discriminating device 43, and only the sound section of the audio signal is converted into cells. Further, only silent sections are detected by the noise detector 44, and background noise in the silent sections is detected by the noise encoder 4.
2. Noise is encoded and transmitted in order to bring out the naturalness of the transmitted voice, but the bit rate is low, so transmitting noise does not significantly reduce line usage efficiency. Therefore, the selector 45 is switched to the noise encoder 42 side when the silence/speech discriminator 43 detects silence.

In addition, only noise (silent frames) is transmitted to the receiving side at an early stage (for example, when starting a communication line connection), and instead of being transmitted thereafter, the receiving side constantly reproduces this noise, and the sending side When it is detected that a change has occurred in the noise, it is possible to make changes such as sending the noise again, or transmitting only the voice without considering unnaturalness and not transmitting the noise. Note that even when no noise is transmitted, it is possible to insert white noise on the receiving side.

FIG. 5 is a detailed block diagram of the silence/speech identification device 43 of the first embodiment of the present invention used in the voice cell conversion device of FIG. Here, the cepstrum is used as the feature parameter of the audio signal. Therefore, an LPC cepstrum detector 51 is connected to the input terminal, and the cepstrum ci (i=1.2, . . . L) of the audio signal is calculated for each frame of a fixed time. The number of parameters is the analysis order, which is 16 in this example, but may be larger since in the present invention the number of feature parameters is reduced after principal component analysis. The cepstrum calculation method is Akin
V, Oppenheim & Rinald W, 5
h1er'DIGITAL SIGANAL PROCE
SSING', PRENTICE) IALL INc
, N month 975) Let C be a vector whose elements are cepstrum c1, C2...cL. The obtained LPC cepstrum vector C is input to the inner product calculator 53. The inner product calculator 53, together with the voiced principal component vector memory 52, calculates the LP of the voiced component of the voice collected under the usage environment of the special device.
The first to third principal component vectors v1, 2, and v3 obtained as a result of performing principal component analysis on the C cepstrum are stored. However, principal component vector Vi (i=1 to 3
) is assumed to be vij(j-1, 2,...'L).

The procedure for determining the voiced principal component vector to be stored in the voiced principal component vector memory 52 is shown as a flowchart in FIG. In step #1, learning audio data generated in advance in the usage environment of the device is collected. In step #2, only voiced data is extracted from all audio data. In step #3, the LPC cepstrum of the voiced data is calculated. In step #4, principal component analysis is performed on the LPC cepstrum. Specifically, an autocorrelation matrix of the LPC cepstrum is calculated. In step #5, the eigenvalues of the matrix are calculated, and in step #6, the eigenvectors corresponding to the eigenvalues with the largest absolute value are sequentially numbered from 1st to Nth.
Let it be a principal component vector. Here, N=3. As a result, the voiced principal component vectors V1, v2, and v3 are obtained.

Returning to FIG. 5, the inner product calculator 53 performs an inner product operation between the cepstrum vector C and the principal component vector Vi as shown in the following equation, and uses the first to third principal component vectors 1, v2, and v3 as coordinate axes. The projection point Q of the LPC cepstrum vector C (=cl, C2, . . . cL)' on the three-dimensional space is determined.

Here, qi is the force direction component of the coordinate axis Vi of the projection point Q.

Also connected is a sound area parameter memory 55 that stores parameters defining an intermediate sound area. 7th sound area
Assuming a rectangular parallelepiped as shown in the figure, the area defining parameters are the lower limit and upper limit of each coordinate axis.
ls V2LN v,,, v3L, vllll. These parameters are also determined in advance by statistically processing the LPC cepstrum of sound components and silent components (including noise) collected under the environment in which the device is used.

The classifier 56 determines whether the frame is silent or has sound, depending on whether or not the projection point Q exists within the sound region of the rectangular parallelepiped shown in FIG. That is, only when V1L≦q1≦v1M and v2L≦q2≦v2,11 and v, L≦q3≦V, l, it is determined that there is a sound, and in other cases, it is determined that there is no sound. The procedure for this determination is shown as a flowchart in FIG.

In this embodiment, the judgment is made depending on whether or not the projection point onto the sound principal component vector space falls within the sound region, but the judgment is made based on the distance between the center of gravity of the sound region and the projection point Q. It is also possible to do so. Let the coordinates of the center of gravity G of the sound area be (gl,
g2, g3), compare the distance expressed by the following formula with a predetermined threshold Th, and if D≦Th, there is a sound,
If D>Th, it can be determined that there is no sound.

D= Σ Ai(qi-gi)2...
...(9) i - where Ai is a weighting coefficient.

According to the silent/sound identification device of the first embodiment of the present invention described above, the L feature parameters are projected into the space consisting of the first to M-th sound principal component vectors, and the projection point is Since the presence or absence of a sound is determined based on whether it exists within the sound region, the number of feature parameters actually used for determination can be reduced, reducing the amount of calculation and making it easier to detect The circuit configuration becomes simple. Furthermore, by projecting into a space consisting of principal component vectors, even if the number of parameters is reduced, the resulting reduction in identification accuracy can be minimized. Furthermore, since the region-based determination is performed, even if the voiced region or silent region exists in a special region on the principal component vector space, it can be determined with high accuracy whether there is sound or no sound.
It is possible to determine whether there is a sound. For example, in equation (9), Ai=1 (i
= 1.2.3), the region where D≦Th is inside the sphere, so when determining based on distance, the shape of the sound region is determined by the definition of distance, and any shape In contrast, in region-based determination, it is possible to set an arbitrary shape.

Note that although the space into which the feature parameters are projected is a space consisting of voiced principal component vectors, it may be a space defined by silent principal component vectors. Furthermore, although a sound area is used for area determination, a silent area may also be used. Although the LPC cepstrum is used as the feature parameter, a spectrum, signal power, number of zero crossings, linear prediction coefficient, autocorrelation function, DFT coefficient, etc. as in the conventional example, or a combination thereof may also be used. Further, specific numerical values such as the number of feature parameters and the number of dimensions of the principal component vector space can be arbitrarily set.

FIG. 9 is a block diagram of a second embodiment of the silence/speech discrimination device according to the present invention. An LPC cepstrum calculator 62 is connected to the input terminal, and the cepstrum ci (il, 2, . . . L) of the input audio signal is calculated for each frame as in the first embodiment. Cepstrum is inner product calculator 63.6
4 is input. A voiced principal component vector memory 65 and a silent principal component vector memory 66 are connected to the inner product calculators 63 and 64, respectively. Inner product calculator 63, 64, voiced principal component vector memory 65, silent principal component vector memory, 6
6 constitutes a feature parameter projection circuit 67. That is, in the second embodiment, the feature parameters (cepstrum) are projected onto the voiced principal component vector space and the silent principal component vector space, respectively. The voiced principal component vector memory 65 stores the first to third principal components obtained by LPC cepstral principal component analysis of the voiced portion of the voice uttered in the environment in which this detection device is used, as in the first embodiment. Stores component vectors. The silent principal component vector memory 66 previously stores first to third principal component vectors obtained by principal component analysis of the LPC cepstrum of the silent portion of speech uttered under the usage environment of this detection device. The inner product calculators 63 and 64, like the inner product calculator 53 of the first embodiment, have the first to third voiced and silent
A projection point Q of the LPC cepstrum vector on a three-dimensional space whose coordinate axis is the principal component vector is determined.

The outputs of the inner product calculators 63 and 64 are detected by the sound detector 68, respectively.
, are supplied to the silence detector 69. The detectors 68 and 69 each have a sound region parameter memory 70 that stores parameters defining a sound region on the sound principal component vector space.
A silent region parameter memory 71 that stores parameters defining a silent region on the silent principal component vector space is connected to the silent region parameter memory 71 . For each parameter, the coordinates of the projection point Q are compared with these parameters, and if the projection point Q exists within the sound area, a detection signal of "1" level is output. Silence detector 6
Similarly, g is “1” if the projection point Q exists within the silent area.
Outputs level detection signal. The outputs of the rain detectors 68,69 are fed to a silence/sound discriminator 72. The discriminator 72 makes the determination as follows.

The output of the sound detector 68 is “1” and the output of the silence detector is “1”.
If the output is “0”, there is a sound, and the output of the sound detector 68 is “0” and the output of the silence detector is “0”.
If the output is “1”, there is no sound, and the output of the sound detector 68 is “1” and the output of the silence detector is “1”.
If the angle is 1°, there is a sound, and if the output of the sound part detector 68 is "0" and the output of the silence detector is "0", there is a sound.

That is, silence is determined only when both the sound detector 68 and the silence detector 69 determine that there is no sound, and in other cases, it is determined that there is sound. The flow of this determination is shown in FIG.

According to the second embodiment, since the determination is made based on the positions of two projection points onto the voiced principal component vector space and the silent principal component vector space, the Even if a voice having a pattern different from the LPC cepstrum pattern of the voiced voice is input and the detection result of the voiced voice detector 68 is silence, the silence detector 6
9 is not determined to be silent, it can ultimately be determined that there is a sound, and failure to detect the presence of sound can be prevented.

Note that the second embodiment can also be modified in the same way as the first embodiment. Furthermore, the determination region of the space consisting of the voiced principal component vector may be set as a silent region, and the determination region of the space consisting of the silent principal component vector may be set as a voiced region.

In the first and second embodiments, silence/voice was identified depending on whether or not the projection point onto the voice/silence principal component vector space was in the voice/silence region. Since there are categories, there may be sounds that cannot be identified in the first and second embodiments.

For example, even if it is pronounced with the - mouth, its characteristic parameters differ depending on each phoneme, whether it is a vowel or a consonant, a male voice or a female voice, or a consonant. Therefore, accuracy can be improved by determining principal component vectors for each of a plurality of categories, determining whether each category is uttered or not, and making a final determination based on the results.

Here, the higher the number of categories, the higher the accuracy, but as in speech recognition, when there is a large number of categories, the scale of the device is large, and the feature parameters of the sound part are classified into a predetermined number of categories, and each category is represented. Find the autocorrelation matrix. Specifically, a method known as the LBG algorithm is used.

A large number of characteristic parameter vectors Xi (i=1 to M) of the voiced parts of the audio collected under the usage environment of the device are obtained, and X
The autocorrelation matrix Ri of i is calculated according to equation (4). By applying the LBG algorithm using a vector whose elements are the row vectors of the matrix Ri as the training vector Ti, a predetermined number of representative vectors Aj (j=1 to N) and division P(Aj) of the problem are obtained. seek. Let the feature parameter vector Xi that created the autocorrelation matrix Ri belonging to the division P(Aj) be a member of the j-th category, and let the matrix Rj created from the elements of the representative vector r] be the representative autocorrelation matrix of the j-th category. shall be.

For the LBG algorithm, Y., Linde, A.
311! Oxnd R, M, Gra7:'An x1g
Oprth bm fo+ Vecto+q u'! I
I wet design', IEEE T+oms
, C0M-28, No l H114-95 (J*nu
xt7.1980).

By the above method, the autocorrelation matrix of the feature parameter vector is classified into a predetermined number of categories, and a representative autocorrelation matrix for each category is determined. LGB
Since the algorithm is used, M autocorrelation matrices R
The error when i (i-1, 2, . . . M) is represented or approximated by Rj (j = 1.2, . . . N) from N (<M) individual representative autocorrelation matrices. The root mean square value is minimized.

Next, principal component vectors for each category are obtained by subjecting the representative autocorrelation matrix Rj of each category to principal component analysis. Also, for each category,
By projecting the feature parameter vector belonging to the category onto the principal component vector space with the principal component vector as the coordinate axis, it is possible to define an identification region on the principal component vector space for determining whether or not it belongs to the category. .

Speech/silence determination is performed by projecting the feature parameter vectors obtained for each category onto the principal component vector space of each category, and comparing the projected points with a predetermined identification area on the vector space.

A block diagram of the third embodiment that makes such a determination is shown in the first example.
Shown in Figure 1. As in the first and second embodiments, a cepstrum calculator 83 is connected to the input terminal 1, and a cepstrum Ci (i=1.2, . . . L) as a feature parameter is calculated for each frame. The cepstrum is input to feature parameter projection circuits 84a to 84j for each category. Here, at t, there are 10 categories. Each feature parameter projection circuit 84 performs principal component analysis for each category.

- As an example, a characteristic parameter projection circuit 84a is shown in FIG. The projection circuit 84a includes a vector memory 92 that stores the principal component vector of category #1, and an inner product calculator 94 that calculates the inner product of the principal component vector of category #1 and the feature parameter vector. In this way, each projection circuit 8
4a to 84j project feature vectors onto principal component vector spaces of categories #1 to #1G, respectively, and obtain projection points.

The outputs of the projection circuits Ha to 84j are sent to the detection circuits 86a to 86.
supplied to j. Each of the detection circuits 86a to 86j detects whether there is a sound or no sound based on the coordinates of the projection point.

As an example, a detection circuit 86a is shown in FIG. Detection circuit 8
Reference numeral 6a includes a category #1 area parameter memory 102 that stores parameters defining a voice area on the principal component vector space of category #1, and a voice detector +04.

Assuming that the principal component vector space is a three-dimensional space, this sound region is also represented by a rectangular parallelepiped as shown in FIG. 7, as explained in the first embodiment, and the parameters consist of lower and upper limits in each coordinate axis. . The detector 104 outputs a "1" level detection signal when the projection point Q exists within the sound region, and otherwise outputs a "0" level detection signal. The output of each detection circuit 86a-86j is supplied to a silence/speech discriminator 88. The discriminator 88 selects at least one
When the outputs of the two detection circuits 86a to 86j are at the "1" level, it is determined that there is a sound. This determination procedure is shown in FIG.

Next, a description will be given of how to classify the feature parameters into a plurality of categories and find the principal component vector of each category and the standard region for silent/speech determination in the principal component vector space of each category.

First, a large number of LPC cepstrum vectors Ci (i-1
, 2, . . . M) and calculate the autocorrelation matrix of Ci according to the following equation. P whose elements are row vectors of matrix Ri
Let the two-dimensional vector be a training vector Ti.

Ti=(ci□　′ ci, C12, -c ・・・・・・(10) ci,, ci2 CI+.

However, ci, ci2. ``'C1p is an element of the LPC cepstrum vector Cj.The training vector Ti is divided into N representative vectors Yj (j-1, 2, . . . N) and the division P( Aj).

Step 1: Initial setting Number of representative vectors N1 Threshold ε of distortion (root mean square value of error between the representative vector and each vector), Initial value of representative vector Ao, ) Training vector Ti (i=1.2, ・M
), set m=o, and set Dl to a large value.

Step 2: Calculation of minimum average distortion Given set of representative vectors Am=(Yj:j=1.
2,...N) such that the minimum average distortion is achieved by dividing P (A
m)=(S i) , i=1.2, ·N is determined by the training vector Ti. That is, the divided area S
For all Ti belonging to i, d(Ti, Yi) x d
(Ti9Tj), j=1.2, . . .N. However, d(Ti.

Yi) is the distortion between Ti and Yi, and can be defined as a squared error as follows.

However, Ti (R) and T i (R) are elements of vector Ti5Yi.

Next, the minimum average distortion due to the division P (Am) is calculated using the following equation.

D,=Dyu ((A,,P(Am))] In the example, N=10, ε-0,01, and M=10000.

The 10 divisions 5i (i=1
~10) become 10 categories, and the LPC cepstrum vector Cj that creates the training vector Tj belonging to Si is said to be a member of the i-th category. Further, the matrix Ri obtained by rearranging the elements of the representative vector Yi becomes the representative autocorrelation matrix of the i-th category.

Step 3: Check for convergence (Dm-+-Dm)/Dm< If so, stop the process and let Am be the final set of representative vectors.

Step 4: Repeat. Set the representative vector set A01 obtained by the current division to A6, set m=m+1, and return to step #2.

The principal component vector of each category can be determined in advance by principal component analysis of K]-. The parameters defining the area of each category on the principal component vector space can be determined in advance for each category by projecting the LPC cepstrum vector belonging to that category onto the principal component vector space of each category.

As explained above, according to the third embodiment, in addition to the effects of the first and second embodiments, feature parameters are projected onto the principal component vector space for each category of sound, and judgment for each category is achieved. Since the results are combined and the final judgment is made, detection accuracy is improved.

Furthermore, since the LBG algorithm is used when classifying categories and finding principal component vectors for each category, it is possible to classify the autocorrelation matrix of M feature parameters into a smaller number of categories, increasing accuracy. can. Note that the third embodiment can also be modified in the same way as the above-mentioned embodiments.

In the above embodiment, when the projection point of a feature parameter is located outside the region for determination in the principal component vector space, the opposite determination is made immediately when the projection point is within that region. In this case, if the determination is made again using other criteria, the determination accuracy will be further improved. FIG. 15 is a block diagram of a fourth embodiment of the silence/speech discrimination device according to the present invention based on such a principle. Similar to the above embodiment,
The audio signal is input to the LPC cepstrum calculator 122.
7125 is connected. The inner product calculator 124 and the voiced principal component vector memory 126 are the feature parameter projection circuit 12g.
Configure. That is, in this embodiment, the feature parameters are projected into the voiced principal component vector space, and the inner product calculator 1
24 determines the coordinates of the projection point Q.

The voiced principal component vector memory 126 stores the first to the voiced principal component vectors obtained as a result of performing principal component analysis on the LPC cepstrum of the voiced part of the voice collected in advance under the usage environment of this detection device.
Stores the third principal component vector.

Also connected are a sound area parameter memo 1J132 that stores parameters that define a sound area, and a silent area parameter memory 134 that stores parameters that define a silent area. The detection circuit 130 determines that there is a sound when the projection point Q exists in the sound area, determines that there is no sound when the projection point Q exists in the silent area, and determines that there is no sound when the projection point Q exists in the silent area. In this case, it is determined that it is indefinite.

The identification result memory 140 is outputted as well as stored in the identification result memory 140. The memory 140 stores identification results of at least three past frames. The output of memory 140 is provided to conditional probability table 138. table 13
8 stores the probability of whether the current frame is silent or has sound according to the identification results of the past three frames, that is, the conditional probability. The conditional probability P is that the identification result D1 of the current frame is 3.
, Di-,,1) When closing i-, P (D i l D,-+, J-z,
Di-i)″P (D・1r Di-z
r Di-i ) However, Di is 1 if the i-th frame has sound, and 0 if it is silent. P (D i D', -,.

Di-z + DI-3) P (DI-1+ D
i-2, DI-3) is based on consecutive 4" frames and 3 frames in which the audio collected in advance under the usage environment of the device is labeled as sound or silent by visual inspection of waveforms and spectra for each frame. The probability is calculated and determined in advance.

The conditional probability that the current frame is speech and silence is supplied to a silence/speech discriminator 136, and the discriminator 136 determines that there is speech if the detection result of the detection circuit 130 is indeterminate. A certain probability is compared with the probability that there is no sound, and the higher probability is used to determine whether there is sound or no sound. The flow of this determination is shown in FIG.

As described above, according to the fourth embodiment, if a positive judgment result is not obtained in the judgment based on the region in the principal component vector space, instead of immediately leading to a negative judgment result, the judgment result obtained from the learning data is This means that the two-stage judgment based on conditional probabilities uses knowledge about speech that it is extremely rare for a line to change from voiced to silent to voiced to silent. Misjudgments of voiced consonants and voiceless consonants with low power are reduced, and the occurrence of omissions and addition of noise at the beginning and end of words is reduced.

This embodiment can also be modified in the same way as the above embodiment.

In the above-described embodiments, whether there is a sound or not is determined simply based on the position of the projection point of the feature parameter, but the accuracy can be further improved if the determination is made based on the temporal change of the projection point. FIG. 17 shows a block diagram of a fifth embodiment of the silent/sound identification device according to the present invention based on such a principle. LPC cepstrum calculator +52 is connected to the input terminal,
Similar to the above embodiment, the cepstrum (i (i=l, 2,
...L) is calculated for each frame. The cepstrum is input to an inner product calculator 154. The voiced principal component vector memory 15 is also connected to the inner product calculator 154, and the feature parameter projection circuit 15B is constituted by the inner product calculator 154 and the voiced principal component vector memory +56. The voiced principal component vector memory 156 stores the first to third principal components obtained by principal component analysis of the LPC cepstrum of the voiced part of the voice uttered under the usage environment of this detection device. . The inner product calculator 154 determines the projection point Q of the LPC cepstrum vector on a three-dimensional space whose coordinate axes are the vectors of the first to third principal components of the voice.

A change vector indicating the change is extracted. Examples of filter +60 are shown in FIGS. 18 and 19. Figure 19 shows the second-order FI
FIG. 2 is a block diagram of an R filter. The filter 160 calculates the projection point vector Q (n) = (q 1 (n), q 2 (n
), q3(n)), filtering is performed on Q(n), and the change vector △Q(n)=(△ql(n)
), △q2(n), △q3(n)). In the case of the filter of FIG. 18, Δqj(n) is expressed as follows.

(16) Here, aj (j=1 to p) is a filter coefficient, and p is the order of the filter. Further, σi is a coefficient for normalizing the dispersion of the filter output, and is expressed as the standard deviation of qi(n) as follows.

(17) On the other hand, in the case of the filter shown in FIG. 19, Δqi(n) is expressed as follows.

△q i (n) = −(x (n) +bl x
(n) +σG bz x (n 2) ) ・
=-(181x (n) = q i (n)
10 a, x (n-1) +a 2 X(
n-2) ・・・・・・
(19) The transfer functions H1(z) and H2(z) of the filters in FIGS. 18 and 19 are respectively expressed as follows.

...... (20) ...... (21) The filter coefficients aj and bj are the transfer functions H1(z) and H2(
z) is set in advance so that it has a bypass characteristic, but
It may be adaptively changed depending on the signal power.

The output of the filter 160 is connected to the matching circuits 162a to 162a.
62j. Each matching circuit 162 calculates the degree of similarity expressed by the Euclidean distance shown in the following equation.

- As an example, a matching circuit 162a is shown in FIG.

Matching circuit 162a is reference pattern table 18
2a and a similarity calculation circuit 1B2a, and the following calculations are performed within this similarity calculation circuit 183.

sm= Σ (△Q 1 r iI I
I ) ) 2i=1 (22) However, R″″ is in the standard pattern table 180a.
=(r1′″tl r2 f+″l, r 3
L s l ) is stored in advance, and (mth standard pattern) is stored in advance.

Note that various other known degrees of similarity may be used as the degree of similarity.

This standard pattern is obtained in advance by the procedure shown in FIG.

First, in step #41, audio data of a portion that can be considered as a sound section of audio uttered under the usage environment of the device is collected as learning data. In step #42, the LPC cepstrum of the learning data is determined frame by frame. In step #43, the cepstrum is projected onto the voiced part principal component vector space, in step #44, a plurality of change vectors representing temporal changes in the projection point are extracted, and in step #45,
Find their centers of gravity and use this as the standard pattern. The reason why a plurality of matching circuits 162 are provided is to calculate the degree of similarity between the change vectors of the projection points of many types of feature parameters and the standard pattern, and to improve the determination accuracy.

The outputs of matching circuits 162a-162j are supplied to silence/speech discriminator 164. The classifier 164 compares the minimum value among the ten similarities with a predetermined threshold, and determines that there is a sound if the minimum similarity is greater than or equal to the threshold, and otherwise determines that there is no sound. The procedure for this determination is shown in FIG.

As described above, according to the fifth embodiment, by determining whether there is a sound or not based on the temporal change of the projection point, there is an effect of preventing false detection due to background noise. This is because although the position of the projection point of the feature parameter may shift due to noise, the temporal change in the projection point is relative and is not easily affected by noise. The fifth embodiment can also be modified in the same way as the above-mentioned embodiments.

FIG. 23 is a block diagram of the sixth embodiment. The output of the LPC cepstrum calculator 202 is supplied to a feature parameter projection circuit 208 comprising an inner product calculator 204 and a voiced principal component vector memory 206. Voiced principal component vector memory
Reference numeral 206 stores first to third principal component vectors obtained by principal component analysis of the LPC cepstrum of a sound part of a voice uttered under the usage environment of this detection device. The inner product calculator 204 determines a projection point Q of the LPC cepstrum vector on a three-dimensional space whose coordinate axes are the first to third principal component vectors of the voice.

The output of the inner product calculator 204 is supplied to the FIR filter 21[1 and the plurality of detection circuits 212a to 212j. Each detection circuit 212 determines whether or not the projection point is in the sound area for each category, and if it is within the area, outputs a "1" level signal, otherwise outputs a "0" level signal. do. An example of the detection circuit 212a is shown in FIG. this is,
a sound area parameter memory 224 for each category;
It consists of a sound detector 226.

The outputs of detection circuits 212a-212j are supplied to plate detector 214. The plate detector 214 tentatively determines that there is a sound if the output of any one detection circuit is at the "1" level, and tentatively determines that there is no sound otherwise.

The output of the plate detector 214 is supplied to the discrepancy detector 216, which determines whether the provisional judgment result of the previous frame and the provisional judgment result of the current frame match, and if they do not match, a "1" level signal FF is sent. otherwise, a signal FF of "0" level is output.

On the other hand, the output of the FIR filter 2+6 is detected by the change detector 218.
is supplied to The change detector 218 calculates the amount of change △ expressed by the following equation using the change vector △Q-value (△q1, △q2, △q3) for each frame of the projection point Q output from the filter 216, and calculates the amount of change. If Δ is greater than or equal to a predetermined threshold, a signal CF of level “1” is output, indicating that there has been a change, and when it is less than that, a signal CF of level “0” is output, indicating that there has been no change.

△=W1△Q1'+W1△ql'+W1△q12...
(23) However, Wl, W2, and W3 are linear weighting coefficients, and Wi is used as the eigenvalue of the autocorrelation matrix of the feature parameter of the voiced part.

Output of plate detector 214, output FF of discrepancy detector 216
, the output CF of the change detector 21g is the silent/sound discriminator 22
0. The silent/sound discriminator 22G first compares the output FF of the mismatch detector 216 and the output CF of the change detector 218, and if they match, the plate detector 214
output as the final judgment result. On the other hand, if the two do not match, the determination is made as follows.

If FF=1 and CF=O, the provisional detection result of the previous frame is output as the final judgment result, and the provisional detection result of the current frame is rewritten to match the result of the previous frame.

If FF=0 and CF=1, the temporary detection result of the current frame is inverted and then output as the final determination result. The flow of this determination is shown in FIG.

In this way, according to the sixth embodiment, since the determination of silence/speech is performed by combining not only the position of the projection point of the feature parameter on the principal component vector space but also the amount of change over time of the projection point, the background This has the effect of reducing erroneous judgments due to noise and improving judgment accuracy.

As explained above, according to the present invention, a large number of feature parameters are calculated from an audio signal, projected into a predetermined principal component vector space (the number of dimensions thereof is smaller than the number of parameters), and a projected point is By identifying whether a phonetic symbol is silent or voiced based on its position, the statistical properties of feature parameters can be used to reduce the number of parameters used for identification.
Moreover, the decrease in identification accuracy due to a decrease in the number of parameters can be suppressed to a minimum. Another advantage is that it is not necessary to optimize the number of parameters and their classification when determining feature parameters.

[Brief explanation of drawings]

Fig. 1 is a block diagram of a conventional silence/speech identification device, Fig. 2 is a diagram showing a spectrum shape to explain the operation of the conventional example, and Fig. 3 is a schematic diagram of silence/speech identification according to the present invention. FIG. 4 is a block diagram of a speech cell generator to which the first embodiment of the silence/speech identification device according to the present invention is applied, and FIG. 5 is a block diagram of the first embodiment of the silence/sound identification device according to the invention. 6 is a block diagram of the embodiment. FIG. 6 is a flowchart showing the procedure for obtaining data stored in the voice principal component vector memory of the first embodiment. FIG. 7 is a standard for identifying silence/speech in the first embodiment. FIG. 8 is a flowchart showing the silence/speech discrimination operation of the first embodiment. FIG. A block diagram of the embodiment FIG. 10 is a flowchart showing the silence/speech discrimination operation of the second embodiment. FIG. 11 is a block diagram of the third embodiment of the silence/speech discrimination device according to the present invention. FIG. 13 is a block diagram of the feature parameter projection circuit of the third embodiment; FIG. 13 is a block diagram of the detection circuit of the third embodiment;
15 is a block diagram of the fourth embodiment of the silence/speech identification device according to the present invention. FIG. Flowchart showing sound identification operation FIG. 17 is a block diagram of the fifth embodiment of the silent/sound identification device according to the present invention, FIG. 18 is a block diagram of the first example of the FIR filter of the fifth embodiment, and FIG. 19 is a block diagram of the second example of the FIR filter of the fifth embodiment, FIG. 20 is a block diagram of the matching circuit of the fifth embodiment, and FIG. 21 is a block diagram of the standard pattern memory of the matching circuit of the fifth embodiment. FIG. 22 is a flowchart showing the procedure for determining the standard pattern; FIG. 22 is a flowchart showing the silent/speech identification operation of the fifth embodiment; FIG. 23 is a block diagram of the sixth embodiment of the silence/sound identification device according to the present invention; FIG. 24 is a block diagram of the matching circuit of the sixth embodiment. FIG. 25 is a flowchart showing the silent/sound discrimination operation of the sixth embodiment. Law: Kenji Oka, Mitsuyuki Matsuyama, Figure 1, Tomo Kuzu, Figure 10, Figure 14, Figure L, J, 2, Figure 3, UT, Figure 5 16 Figure 8 Figure 9 Figure 22 Figure 24 Figure 25

Claims (2)

[Claims] (1) The principal component vectors of the sound part and the silent part are calculated in advance for the audio data collected under a predetermined environment, and the feature parameters of the frame to be detected are set in the principal component vector space of the sound part or the silent part. A voice detection method is characterized in that the speech is projected onto the principal component vector space of the area, and the presence/absence of speech is determined based on the position of the projection point and the amount of change over time. (2) Calculate the principal component vectors of the voiced part and silent part in advance for the audio data collected under a predetermined environment, and set the feature parameters of the frame to be detected in the principal component vector space of the voiced part or the silent part. A voice detection method is characterized in that the speech is projected onto the principal component vector space of the area, and the presence or absence of speech is determined based on the amount of change over time of the projection point.