JPH02203397A - Voice/voiceless part detection system - Google Patents

Abstract

PURPOSE:To improve the detection accuracy by deciding which category a feature parameter belongs to be using the position of the projection point of the feature parameter in a main component vector space of each category and a specific area predetermined in the main vector space. CONSTITUTION:Means 150 - 170 decide whether or not the feature parameter belongs to categories by using the positions of projection points by projecting means 120 - 140 for the feature parameter in the main component vector spaces of the categories and specific areas predetermined in the main component vector spaces. Then the feature parameter is projected in the main component vector spaces and then a voiced/voiceless decision is made. Consequently, even when the number of parameters used for the voiced/voiceless decision making is decreased, the loss of information that the original feature parameter has is least, the voiced/voiceless sound detection accuracy is high, and the absence of a voice and the addition of noises due to an error in the voiced/voiceless decision making are reduced.

Description

[Detailed Description of the Invention] [Objective of the Invention] (Field of Industrial Application) This invention converts and transmits only the sound portion of audio into cells.
T M (Asynchronous Tranous
(Mode) Recording system that records only the voiced part of communication and sound/silence detection method of voice signal, which is the basic technology of voice recognition (conventional technology) ATM communication that records only the voiced part of voice and transmits it ,
f Voice recognition, recording people who record only the sound part,
A voice/voice section that detects the voice interval or the start and end of the voice.
Silence detection is the most basic and important process. Sound/
If silence detection is not performed accurately, audio may be interrupted or speech recognition errors may increase. Special numbers are considered to be the key to effective use of lines in the current ATM communications era.

JP-A-2006-101010 is a conventional example of a voice/silence detection method that can reduce the dropout of initial consonants with a low level even when the ambient noise level is high, without depending on the input signal level according to the signal input condition. 1986-200300 ``Voice start and end detection device'' is known.

This conventional method will be explained below.

FIG. 6 is a block diagram of the start and end detection device described in the above publication. In FIG. 6, 600 is an energy extractor, which is composed of a rectifying and smoothing circuit, and extracts the signal power once per frame. 610 is a spectral shape extractor, which is a low frequency (250 to 600 Hz).

Mid range (600-1500Hz), high range (1500-40
00Hz) 3111JA band pass filter details and rectification smoothing circuit. ! The power of each frame in the r band is used as spectrum information.

Energy extractor 600 and spectral shape extractor 610
This constitutes a feature amount extraction section 620. 630 is a multiplexer for inputting the signal power from 600 and the band filter power from 610 to the sound/non-sound determining section 640 in a time-division manner.

Reference numeral 640 denotes a voice/non-voice determination unit for determining whether there is a silence, an unvoiced sound, or a voiced sound. 650 and 660 are a threshold memory and a standard pattern memory, and are a sound/non-sound determination unit 640
Constant values used in are stored. [[Memory 65
Two power thresholds & and Et are stored in 0f. Additionally, the a* quasi-pattern memory 660 contains silent and
Linear discriminant function for determining voiceless sounds and silent/voiced sounds
Coefficients of a linear discriminant function for discrimination are stored.

These two threshold values E, , E, and the coefficients of the linear discriminant function of 2 are obtained in advance by statistical processing of voice data uttered under the environment to be used, and are stored. 670 is the starting point.
This is an end candidate production section, which detects start and end candidates of audio based on the duration of the sound/no sound output for each frame sent from the sound/no sound determination section. 680 is a start/end determining unit that determines the final start/end.

To explain the actual process of detecting the start and end of audio configured as described above, first, the signal representing the audio input from a microphone etc. is logarithmically calculated every frame. Power LPW and logarithmic band power I, P i
(t = 1 to 3) (This is converted. Sound/silence determining unit 6
40 is a memory 650 for these four parameters and a threshold value.
and standard pattern memory 660f are stored. threshold 1
111Et*Et and the coefficient 2 of the two linear discriminant functions are used to determine whether there are i or itξ input frames.

This sound/non-sound judgment starts with two energy thresholds E,,E! A comparison of 1 and the logarithmic power LPW is performed as follows.

If LP W ) E r then there is sound LPW<Et
Then, silence E, LPW<;:E, then indefinite indefinite 'a e + ko, furthermore, logarithmic band power LP
Using i (1=1 to 3) and the coefficients of the two linear discriminant functions stored in 660, the discriminant function value FX of Equation 1 (1) is calculated, and voice presence/absence is determined from the FX.

However, At is a coefficient of the discriminant function stored in 660f, and LPi is a standard pattern stored in 660.

A18 and LPl in Equation (1) are obtained in advance by statistically processing silent, unvoiced, and voiced sounds of audio data uttered under the usage environment. The FXQ value is set so that it takes a negative value when the input is silent, and a positive value when the input is voiceless or voiced. To determine whether there is sound or no sound based on the spectral shape, calculate two linear discriminant functions: silent/voiced and silent/voiced, and if either one has a positive value, it is determined that there is a sound, and if both values are for soldiers, then it is determined that there is no sound. It is about making a judgment. This method utilizes the differences in the spectral shapes of silence, voiceless hakama, and voiced consonants, so it has the characteristic of reducing the omission of voiceless consonants and voiced consonants with low energy.

However, with this method, there are few parameters that represent the spectral shape, and there is no theoretical basis for selecting parameters, so there are cases where voice/no-voice determination is incorrect and audio dropout or noise is unavoidable. There is. The parameters of this method are reduced (250-600Hz)
, midrange ([)0-1500Hz).

It is the logarithmic power of the output of three band filters in the high frequency range (1500 to 4000 Hz). For example, as shown in %6-6, the spectrum of unvoiced sound is (a) and the spectrum of noise is (b). In case 5, the value of the linear discriminant function calculated by equation (1) will be the same even though the spectra of the two are greatly different, and
! 41 constant (however, At=1). As a result, dropout of audio or addition of noise may be unavoidable.

This problem is caused by the small number of parameters and inappropriate selection of bandpass filters. In addition, since there is no theoretical basis for the parameter selection method, the selection of parameters, that is, the setting of the band of the bandpass filter, requires a lot of effort and effort to set the parameters by trial and error. The problem is that it is not necessarily appropriate. If the number of bandpass filters is increased and the number of parameters is increased, errors in speech/non-speech determination can be improved, but the total ambiguity of the discriminant function for speech/non-speech determination will increase. Furthermore, the effort required to set the parameters becomes enormous. Although the technique disclosed in the above-mentioned publication describes that the Mahalanobis metric can be used instead of the linear discriminant function of Equation 1 (1), using the Mahalanobis metric increases the amount of calculation by 11.

(The problem that the invention is trying to solve) As mentioned above, conventional O presence/silence detection is removed.

In order to reduce the number of calculations, [the number of parameters was reduced by 3* &, there is a problem that the presence/absence determination P may be incorrect and the omission of vocalizations or the addition of noise may be unavoidable. In addition, the conventional method has the problem that it requires a lot of effort because there is no idiomatic selection basis for determining the parameter selection order.

Therefore, the present invention has been made to solve these problems, and its purpose is to provide a voice/silence detection method that has high voice/silence detection accuracy and is less likely to drop out voices or add noise. There is a particular thing.

[Beak formation of invention]

(Means for Solving the Problems) In order to solve the above-mentioned problems, the present invention provides means for determining characteristic parameters such as signal power and LPC factor a representing the characteristics of acoustic signals such as audio signals, and After finding the characteristic parameters of the audio signal including the miscellaneous TF-E that is assumed when using the communication port and processing technology of a device, etc., and classifying the autocorrelation of these characteristic parameters into multiple categories, each Principal component vector 7i of each category obtained by principal component analysis of category! , a means for shaping the special enemy parameters for each frame on the space, a means for projecting the projection point of the feature parameter on the flea principal component vector space for each category, and a specific area predetermined on the principal component vector 2 space. means for determining whether a feature parameter belongs to the category, and means for combining the determination results for each category to determine whether or not there is a sound.

(Operation) First, characteristic parameters of an acoustic signal such as a voice signal are determined. Next, consider reducing the number of parameters to less than the binary feature parameter after converting that parameter to another parameter by f +A. Figure 4 shows this concept. Figure 4 (here, the five original feature parameters r, xl (i=1.2s
-+L)t! :, and xl. Let X be the vector to be used as the vector. ′ is an orthogonal transformation, and the transformation matrix is A. Feature parameter after conversion i y 1 (i=1.2...e
L ) -y The vector whose elements are Y, leaving N parameters yj (j = 1, 2..., N), f
Let Y be a feature parameter vector with iFl)(L-N) zeros.

However, it is NIL.

At this time, the error vector e caused by the reduction in the number of parameters l is described as the difference between the original 1-lose parameter vector X and the inverse transformation of Y as shown in the following equation.

6=X-A-IY =A-'(Y-Y) L2) By performing a transformation that minimizes the root mean square value of this error fr'' ==E (ete), it is possible to reduce the number of feature parameters. The error is minimized, where t is the transpose of the matrix,
E is the expected value. The transformation to minimize rl is the matrix A, which is the eigenvector of the autocorrelation matrix of xl.
It is known that there is a transformation that depends on t, that is, a KL transformation. In addition, the eigenvectors are the same as the principal component vectors obtained by principal component analysis of xl (the eigenvectors corresponding to the one with the largest eigencompensation are 1st, 2nd, 3rd, etc.)
...corresponds to the principal component vector.

The operation of reducing the number of parameters after KLf of L feature parameters . Therefore.

By projecting the feature parameters onto the principal component vectors, we can express the feature parameters of one element with fewer parameter dimensions, while minimizing the loss of information contained in the original feature parameters. The number of feature parameters can be reduced.

The characteristic 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.

Speech/silence determination is performed by comparing the projected point of the feature/sound meter on the principal component vector space and the predetermined voice and silence regions on the principal component vector space, using the basic property of i-cohoko. It can be done with high precision.

In order to more accurately determine the presence/absence of sound, the present invention includes the following steps:
The feature parameters of the voiced part are classified into multiple categories, the voiced/silenced judgment is made for each category, and the results are integrated to make the final voiced/silenced judgment. Even if it is said that there is a sound at -, the characteristic parameter is 1, for example.

This is because vowels and consonants, male and female voices, and even consonants differ depending on the beginning of each sound. Therefore, if the percent speech is classified into as many categories as possible and the presence/absence of speech is determined within each category, the accuracy of speech/non-speech detection will be higher. However, if the number of categories is increased by more than Cζ for recognition, there is a problem that the scale of the apparatus becomes large, so it is necessary to narrow down the number of categories to an appropriate number.

Here, we will explain how to narrow down the number of categories and how to narrow down the number of categories.
xl (1=1, 2,...,M), each x
The root mean square error of Equation (2) for 1 is 11 [E d small]. The transformation is calculated by the following equation. .

It is an element of vector Xt.

From equation (3), it can be seen that the autocorrelation matrix R is the average of the autocorrelation matrix R1 with each feature parameter vector expressed by a linear equation, I, in a dimensional space, that is, the center of gravity.

The autocorrelation matrix Rt(
When considering that t-4, 2, ......, M) is represented by one autocorrelation column and row Tary R, the autocorrelation matrix R obtained by equation 1 (3) is a linear equation. The defined Ri and R2 minimize the root mean square error E.

However, xil , xil , turn times ·, xiL are characteristic parameters. Therefore, by partially differentiating the above equation with respect to R(k, j), O
This is because π(k, j) and 11(lc, j) are respectively self-phase 1 column row 91J R and R1. (k, j) elements.

Based on this way of thinking, we have created a division of existence! l? The micro/parameters are classified into a predetermined number of categories IJ-1, and a representative autocorrelation matrix for each category is determined. Specifically, the hand f known as the LBG algorithm
Use f.

First, if the characteristic parameter vector of the vocal part of the voice is collected in advance in a convenient environment such as a telephone, X1 (1=1°2, . . .
..., M), xtco autocorrelation line flJ
Calculate R1 according to equation (4). Next, the training vector t
By applying the LBG algorithm as 1, a predetermined number of representative vectors ^j(j-1, 2,...
..., N) and the division P(Aj). If the feature parameter vector Xl that created the autocorrelation matrix R1 belonging to the obtained division P(Aj) is a member of the j-th category, then 'g! Let matrix Rj derived from the elements be a representative autocorrelation matrix of the j-th category. For the LBG algorithm, Y,
Linde, ^, Buzo and R, M.

Gray: “An algorithm for
Vector quantizer design IEEE Troms, C0M-28,
No, lpp.

84-95 (January, 1980).

(Sentence #2) By the above method, when the feature parameter vector is divided into a plurality of predetermined categories, the representative autocorrelation yIj of each category is found. Since this method uses the LBG algorithm, 1M autocorrelation matrices R1 (i=1.2°...
M) from 8 (kuM)Q representative autocorrelation rows 11 [(j
Engineering 1.2. ..., N), the root mean square value of the error when approximated or approximated is minimized.

Next, representative autocorrelation matrix IR for each category
Principal component vectors for each category are determined by principal component analysis of j. In addition, for each category, by projecting the feature parameter vector that belongs to the category on the principal component vector space with the principal component vector as the coordinate axis, we can create a region for determining whether it belongs to category #C or not. predetermined on the component vector space.

Speech/silence determination is performed by projecting the feature parameter vector obtained for each frame onto the principal component vector of each category, and comparing the projected point with a predetermined area to determine whether it belongs to category 1 or not. After examining each category, the results of each category are summarized.

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

First, the voice cell conversion trap 11f used in ATM communication using the present invention will be explained with reference to FIG. This device utilizes lines effectively and transmits at high speed.One side of the input audio signal is sent to the audio encoder
2 (encoded here.

The other encodes the noise signal in a noise encoder 703. And these encoded signals are converted into cells [17
05, it is converted into cells and transmitted. When encoding audio, the presence/absence production device 701 detects a sound part and a non-speech part, and controls the switch Ij to convert only the sound part into cells. Also, regarding noise, the voiced noise detector 7
At step 04, only voiced noise is detected and encoded.

Noise is provided to give naturalness to the voice, and when a voice other than voice, that is, silence is detected, the switch 704 is switched to the noise encoder 703 side and the transmission is performed.4Transmission method in this block diagram ( An example of a method for transmitting a sound part and a silent part (including noise) by applying different encoding, or a different video tray) (24, 8
A possible method is to transmit the data at a high speed (Kbps).

Also, only noise (silent frame), which is not shown in this block diagram, is transmitted to the receiving side at an early stage (for example, at the time of connection), and this noise is always reproduced on the receiving side when a predetermined change is detected. Possible methods include a method in which this noise is retransmitted only when the noise is detected, and a method in which only voice is transmitted and no silence is transmitted at all. In the case of (2), the noise may be reproduced using white noise or the like that is present on the receiving side.

We will discuss in detail the sound and silent effects used in the above system.

FIG. 1 is a block diagram of a speech/silence detector according to an embodiment of the present invention. In FIG. 1, 5110 is an LPC cepstrum extraction circuit, which extracts the LPC cepstrum C1 (t=i, z, . . . ) of the signal input from the input terminal 100.
...', P) f is calculated for each frame (10 ms) using a known method. However, P is the analysis order, and is assumed to be P=16, for example. The calculation method of the LPC cepstrum is described, for example, in Sadahiro Furui's ``Digital Speech Processing'' (Tokai University Press, 1985). 120
is the feature parameter projection circuit 41:1, and the LPC cepstrum vector C=(C1, C1
...Cp) Category IJ-4 for which t is determined in advance
: Project onto the principal component vector space of 1. Similarly, this 130 converts the LPC cepstrum vector C into a predetermined category 4! This is a feature parameter projection circuit 4I-2 that projects onto the principal component vector space of "2".In this embodiment, the number of categories is 10, and a feature parameter projection circuit is provided for each category.

Figure 2 shows the feature parameter projection circuit shown in Figure 1≠11
20 is a block diagram showing a configuration example of the inner product calculation circuit 2.
10 and a category≠1 principal component vector memory 200. Feature parameter projection circuit ヰ2~-jIP1
(The same applies to H getting stuck.

Category principal component vector memory This is the first to third three principal component vectors V of category≠1 determined in advance.
,,V,,v, are stored. The method of classifying categories and how to obtain principal component vectors will be described later.

The inner product calculation circuit 210 performs the inner product calculation of the LPC cepstrum vector C, which is the output of the circuit 110, and the principal component vectors 2, ˜V, according to the following equation.

Vs * Vs Project C onto the three-dimensional principal component vector 9 with the f coordinate axis, and calculate the coordinates of the projection point Q, Qt, Qm-Q
Find i.

However, Vij is a reasonable value of principal component v1. The coordinates Ql(
i=1.2.3) are category judgment circuits 150, 160
．． 170, where it is determined for each category whether the LPC cepstrum extracted from the input signal belongs to that category. FIG. 3 is a block diagram showing an example of the construction of the category $1 judgment circuit shown in FIG. Category Studies 2-101! The same applies to the IJ constant circuit. Category IJ −+ 1~:lI:l O? The fJ area defining parameter memory stores parameters that define the area of each category on the principal component vector space of each category. If the area of each category is a rectangular parallelepiped as shown in Figure 4, the parameters that define the area are V,
11/Ih, Vt j, Vt h, Vs 1.
'ush (!: f! 6゜These parameters are
The method is determined in advance, and the details will be explained later. The determination circuit 230 in FIG. 3 determines that the projection point Q is @4
It is determined whether the input feature parameter belongs to the category =l:ll depending on whether it exists in the area of the diagram. Namely.

1/”'i l <Q t <:Vt h katsu IJ*
l <Q t 2===゛1. )t h?”')Vs
When l<Q s<1)s h■, it is determined that it belongs to the category≠1, and @1" is output; otherwise, 10" is output.

Returning to FIG. 1, the utterance/silence determining circuit 180 performs utterance/silence determination by integrating the outputs of the category IJ -=$1 determining circuit to the category≠10q constant circuit.

Specifically, category≠1 judgment circuit ~ category 41
:l The output of the O determination circuit is ORed, and if the result is "1", it is determined that there is a sound, and if the result is "O", it is determined that there is no sound.

The above is an explanation of the operation of the f/silence detector according to one embodiment of the present invention. Below, a method for classifying feature parameters into a plurality of categories and determining the principal component vector of each category and the region in which each category exists on the principal component vector space will be described.

First, the LPC cepstrum vector C1 (1=1.2...
..., M), and calculate the autocorrelation matrix R1 of C1 according to formula (9) I.0-order fc1 matrix Ri
P with O row vector elements! Let the dimensional vector be the training vector tl.

(So J t 1 = (Ci, t c i, Cil...
・・C1, C1pCil CtlCi♂・・・・・・・
・・・ctlcip・・・・・・・・・・・・・・・C1
However, ci,, C11..., Cip
The beam is an element of the PC cepstrum vector C1.

Also, t indicates the transpose of the matrix, the training vector t
1 applies the LBG algorithm and calculates N representatives, vector yj (j = 1.2...
, N) and the division P(Aj).

0th step: Initial setting Number of representative vectors
, ...-, M) and set m=o, l), = .

First step: Result of the given representative vector Am=(
yj, j=1. ...1, N), such that the minimum average distortion is achieved by dividing P (Am) = lSi), i = 1.2
．． -, N are determined using the training vector ti.

Namely.

All tits belonging to the divided area Sif are added, d(
ti, yl)<lti, yj), j=1.2. ...
・・・・・・Rub E so that it becomes N. However, d(ti
, yl) is the distortion between ti and yl, and can be defined as a squared error like a first order.

t 1(R), y l('EQl'it i, y
iO is ammonium.

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

Dm=Dm[(Am, P(Am))) L creating the toning vector tj belonging to Sl
It follows that the PC cepstrum vector Cj is a member of the first category. Also.

The matrix R1 obtained by changing the number of elements of the representative vector 71 (1=1.2...,N) becomes the Daiken autocorrelation matrix of the first category.

Second step: If convergence check (Dm-+-Dm)/Dm(grl) is reached, the process is stopped and Am is set as the final collection of representative vectors.

Third step: Repeat the representative vector assembly A m obtained by the current division.
10 + f A rn and return to the first step with m = m + sign.

In this embodiment, the initial setting is 2, N=10, and ε=
0. Let Ol, m=10000.

The 10 divisions 51 (1=1
．． 2. ......, N) becomes 10 categories, and the principal component vector of each category is calculated by principal component analysis 1 of R1.
It can be obtained in advance from this. In addition, the main component Pegtor 2
The parameters that define the area of each category on the !
It can be predetermined for each category by projecting the LPc vector to which the category belongs onto the principal component vector space of the category.

In the above embodiment, the parameters used for the sound/tuskless determination are #Qt-Qs (D3) for each category, but the number can be set arbitrarily. Even if the number of categories is reduced by the number of feature parameters in one element, the error is minimal.Also, in this example, the number of categories is 210, but the number can be set freely. The error due to classification into a finite number of categories is minimal.

Furthermore, in this embodiment, it is determined whether the input feature parameter belongs to each category 17-1 by determining whether the projection point on the principal component vector space falls within a specific area determined for each category. This is done using the distance between the center of gravity of the area and the projection point. For example, when the center of gravity of the region is V = ("LF" + - "vtVs), the distance fiD is defined by the following formula, and D is compared with a predetermined threshold Th. From this, if D < T h rl, the category , and if D>Th, it can also be determined that it does not belong.

1=1 However, al is a weighting coefficient.

In the conventional voice/silence detection method, a determination based on Takashima is used, and determination based on an area has not been used in the past. The determination based on the area Cζ allows the presence or absence of voice to be determined even when voice or silence is distributed in a special area on the principal component vector space, so it has the effect of improving the accuracy of voice/silence detection. For example, equation (14) where l = 1 (1 =
In the case of 1.2.3) and 2, the region where D<:Th is inside the sphere.In the judgment based on the distance 1lf16c, the shape of the voiced region is determined by the definition of the distance, and any shape can be formed. cannot be set, whereas in region-based determination, any shape can be set.

As for the characteristic parameters of the input signal, in addition to the LPC cepstrum, signal power zero crossing a, LPC coefficients, autocorrelation coefficients, DFT coefficients, and combinations thereof can be used.

As described above, the present invention performs voice/silence determination after projecting the feature parameters onto the principal component vector space, so even if the number of parameters used for voice/silence determination is reduced by 2, the The loss of information of the characteristic parameters is the smallest, the presence/absence of tm detection accuracy is high, and the dropout of voice and the addition of noise due to failure of speech/non-speech detection can be reduced. In selecting parameters to be used for voice/silence determination, first, many feature parameters are determined, and the parameters and the second principal component vector are selected in order from the first principal component vector with the largest eigenvalue. Third...There is a reasonable criterion that the loss of information in the original feature parameters will be minimized if the parameters obtained by performing inner product calculation on the principal component vectors are used.

This has the effect of making it easy to set the parameters used.

Furthermore, in the present invention, the feature parameters for each frame are projected onto the principal component vector space of each category obtained by ffl, which classifies the autocorrelation matrix of the feature parameters into a plurality of categories, and the principal component analysis number of each category. Since the presence/absence determination is made for each category and the results are combined to detect the presence/absence of speech, there is an effect that the accuracy of detecting the presence/absence of speech is improved. Moreover, since the LBG algorithm is used to calculate the category a and the principal component vector for each category, the error caused by classification into fewer categories than the autocorrelation matrix 9M of M feature parameters can be minimized. This has the effect that it is possible to improve the accuracy of sound/non-sound detection.

〔Effect of the invention〕

As described above, according to the present invention, the amount of calculation is small - in spite of this,
This has the effect that it is possible to accurately determine whether there is a sound or no sound, and the reliability of the system is also improved.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a speech/silence detector according to an embodiment of the present invention, and FIG. FIG. 3 is a block diagram showing an example of the configuration of the category IJ-4:1 determination circuit shown in FIG. FIG. 5 is a diagram showing the concept of reducing the number of feature parameters used in the detailed explanation of the present invention. FIG. 6 is a diagram showing the conventional f/silence detection tt[l Block diagram: FIG. 7 is a diagram showing an example of spectra determined to have the same spectral shape by a conventional f/#f detection device; FIG. 8 is a block diagram of the voice cell f-1 according to the present invention. . 100... Input terminal, 110... LPC cepstrum extraction circuit, 120, 130, 140... Feature parameter projection time if! , 150.160, 170...Category judgment circuit, 180...Sound/absence f judgment circuit, 2
00... Category principal component vector memory, 210...
- Inner product calculation circuit% 220...Category area defining parameter memory, 230...Judgment circuit, 600...
Energy extraction section % 610... Spectrum shape extraction section, 620... Feature amount extraction section, 630... Multiplexer, 640... Sound/non-sound determination section, 650...
Threshold value memo IJ, 660...Standard pattern memo IJ, 6
70...Starting end-terminus candidate detection unit, 680...Starting end-
Termination determining section.

Claims (3)

[Claims] (1) A means for obtaining a feature parameter representing the feature of an acoustic signal of an audio signal, and an autocorrelation matrix of the feature parameter of the audio signal including preset noise, for dividing the feature parameter obtained by this means into multiple categories. A means for projecting onto the principal component vector space of each of the classified categories, a projection point position of the feature parameter on the principal component vector space for each category projected by this means, and A means for determining whether the feature parameter belongs to this category using a specific area, and a means for determining whether there is a sound or no sound using the determination result for each category determined by this means. This is a sound/silence detection method that is characterized by: (2) The voice/silence detection method according to claim 1, wherein the means for determining the feature parameters determines the feature parameters using signal power, LPC coefficients, etc. representing the characteristics of the acoustic signal. (3) The principal component vector space of the feature parameters of the voiced part or the silent part set in advance is obtained by principal component analysis of the feature parameters of the device used. Silence detection method.