JPH0458297A - Sound detecting device - Google Patents

Abstract

PURPOSE:To obtain a satisfactory detection rate without being affected by an input level even in the environment of large background noise by discriminating an input signal for each frame unit by using a noise standard pattern and a sound standard pattern. CONSTITUTION:A feature vector calculator 1 calculates a predictive linear coefficient by using a Durbin method or the like for the unit of a frame. A feature vector X(n), which is estimated as noise by a noise block estimating device 2, is transferred to a noise standard pattern preparing device 7. The device 7 prepares the noise standard pattern. Based on the feature vector X(n) outputted from the feature vector calculator 1, a discriminating device 10 discriminates a sound block. By using the feature vector estimated as noise, the noise standard pattern is prepared and based on this noise standard pattern and the sound standard pattern, the discrimination of sound/noise is executed. Thus, the satisfactory detection rate can be obtained without being affected by the input level even in the environment of large background noise.

Description

[Detailed description of the invention] [Object of the invention] (Industrial application field) The present invention is directed to an ATM (Asynchronous Tr).
ansf'er Hade) communication, DSI (D
digital 5peech Interplati.

n), relates to a sound detection device used for detecting a sound interval in a voice signal in the field of packet communication, voice recognition, etc.

(Prior Art) Speech detection devices play an important role in the fields of communication and speech recognition, such as improving communication efficiency and recognition rate.

FIG. 9 is a diagram showing the configuration of a conventional sound detection device. First, a feature vector calculation device 23 calculates feature vectors such as the power of an input signal, the number of zero point differences, an autocorrelation function, and a spectrum. Thereafter, the matching device 24 measures the distance between this feature vector and the speech standard pattern and the noise standard pattern. For example, if the distance between the feature vector and the standard speech pattern is shorter, it is determined to be speech, and conversely, if the distance between the feature vector and the noise standard pattern is shorter, it is determined to be noise.

Further, FIG. 10 is a diagram showing another example of a conventional sound detection device, in which a power calculation device 25 calculates the average power P (n) of an input frame. In addition, the threshold value updating device 2
6 is the threshold value T(n) for determination, for example, if P (n) < T (n) − P (n)
(a −1), then T (nil) −P (n+1
) X a if P (n)≧T (n) −P (n)
If X (a −1), then T (n+1) −
Update as P (n+1) X 7.

Here, α and γ are constants.

Then, the threshold comparison device 27 compares the input frame with the input frame if P (n)≧T (n), then if P (n)
< T (n), it is identified as noise.

Note that the threshold value updating device 26 updates the threshold value as follows: If P (n) < T (n) - α, then T (n
+1) = P (n+1) + aIf P (n)≧
If T (n) − α, then T (nil) −P (n+
1) It may be set to +7.

By the way, the power of most vowels generally exceeds the background noise power, but the power of consonants is often lower than the background noise power. Therefore, in environments with large background noise,
Even in the consonant interval, noise features appear significantly in the feature vector.

However, in the conventional speech presence detection device described above, the feature vectors affected by background noise are used as they are for determination, so if the background noise is large, consonant detection errors may occur. . Further, in conventional speech presence detection devices that use only power as a speech feature vector, there is a problem in that speech is often incorrectly determined to be noise when the input level is low. However, such a situation has caused a deterioration in sound quality in the field of communications, and a decrease in recognition rate in the field of speech recognition.

(Problem to be Solved by the Invention) In this way, in the conventional voice detection device, the feature vectors affected by the background noise are directly used for determination, so when the background noise is large, consonant detection is difficult. There is a problem with errors occurring.

Further, in conventional speech presence detection devices that use only power as a speech feature vector, there is a problem in that speech is often incorrectly determined to be noise when the input level is low.

The present invention was made based on the above circumstances, and an object of the present invention is to provide a sound detection device that can obtain a good detection rate even in an environment with large background noise without being affected by the input level. It is said that

[Structure of the Invention] (Means for Solving the Problem) In order to solve the above-mentioned problem, the present invention provides a feature vector calculation means that divides an input signal into frame units and calculates a feature vector of the input signal for each frame. a voiced interval estimation means for estimating a voiced interval of this feature vector based on the feature vector calculated by this feature vector calculation means; and a feature of a frame estimated to be not a voiced interval by this voiced interval estimation means. Calculated by the feature vector calculation means using a noise standard pattern creation means that creates a noise standard pattern based on the vector, a noise standard pattern created by the noise standard pattern creation means, and a voice standard pattern created in advance. and determining means for determining whether the input signal for each frame is speech or noise from the determined feature vector.

A second invention also provides a feature vector calculation means for calculating a feature vector of an input signal in units of frames from an input signal, and a feature vector calculated by the feature vector calculation means in which fluctuations in the input signal are removed. a feature vector converting means for converting into a conversion vector; a sound interval estimating means for estimating a sound interval for each frame based on the feature vector; and a conversion vector estimated to be not a sound interval by the sound interval estimating means. A noise standard pattern creating means creates a noise standard pattern based on the noise standard pattern creating means, and a noise standard pattern created by this noise standard pattern creating means and a pre-created audio standard pattern are used to convert the input signal of each frame into audio. and a determination means for determining whether it is noise or noise.

(Function) In the present invention, a noise standard pattern is created using features/conversion vectors estimated as noise, and this noise standard pattern and a pre-created audio standard pattern are used to input each frame. Since it is determined whether the signal is voice or noise, a good detection rate can be obtained regardless of the input level even in environments with large background noise.

(Example) Hereinafter, details of embodiments of the present invention will be explained based on the drawings.

FIG. 1 is a block diagram showing a schematic configuration of a sound presence detection device according to an embodiment of the present invention.

In the following, we will analyze the input signal frame by frame and determine whether it is speech or noise. For example, the input signal is sampled at 8 KHz, and each 160 samples is made into one frame. However, the frame length and analysis cycle do not always need to be constant lengths.

The feature vector calculation device 1 performs Durbi on a frame-by-frame basis.
Calculate linear prediction coefficients using the n method or the like. Here, linear prediction coefficients may be converted to calculate PARCOR coefficients, LPC cepstrum, mel cepstrum, etc., and may be used as feature vectors. Also, power, autocorrelation function, number of zero crossings, etc. may be calculated.

The frame for which it is currently being determined whether it is speech or noise will hereinafter be referred to as an input frame. Further, the feature vector of the input frame obtained by the feature vector calculation device is assumed to be X(n). n is the sequential number of the frame.

The feature vector is a p-dimensional vector, X (n) = (x 1(n), x2(nl...xp
(n) ).

The noise interval estimating device 2 includes, for example, a power calculation device 3, a threshold updating device 4, and a threshold comparing device 5, as shown in FIG.

Here, the power calculation device 3 calculates the average power P of the input frame.
Calculate (n). Further, the threshold updating device 4 updates the threshold T (n) using the following equation.

If P (n) < T (n) − P (n) X
(a −1:), then T (n+1) = P (n
il) x a if P (n)≧T(n) P(n)
x Ca −1), then T (n+1) −P (
nil) x 7 where α and γ are constants.

The threshold comparator 5 compares the input frame with P (n)
If ≧T (n), then voice, if P (n) < T (n
), the output switch 6 that identifies it as noise transfers the feature vector X (n) estimated to be noise by the noise section estimation device 2 to the noise standard pattern creation device 7 .

The noise standard pattern creation device 7 creates a noise standard pattern, and for example, as shown in FIG.
and an average/covariance matrix calculation device 9. The buffer 8 stores the feature vectors X (n) estimated to be noise by the noise interval estimating device 2. Then, the mean/covariance matrix calculation device 9 creates a noise standard pattern based on the feature vectors accumulated in the buffer 8.

The determination device 10 determines a voiced section based on the feature vector χ(n) output from the feature vector calculation device 1. For example, as shown in FIG.
1. Consists of a voice standard pattern storage device 12 and a noise stack pattern storage device 13 that stores the noise standard patterns created by the noise standard pattern creation device 7.

The matching device 11 measures the distance between the voice standard pattern and the noise standard pattern and the feature vector If it matches the noise standard pattern, it is determined to be noise.

Specifically, first, each standard pattern (Σj, μ
i) Measure the distance to (i=1.2...M+1).

Di (f) = (X-uI) ' If -'
(X-μm)+In Σ l X assumes that χ belongs to wi, i -5in Di (X). If wl belongs to voice, the frame is determined to be voice, and if wi belongs to noise, the frame is determined to be noise.

Here, the voice standard patterns stored in the voice standard pattern storage device 12 can be defined as follows.

As can be seen from the configuration of the matching device 11, the standard pattern is the mean value vector μ and covariance matrix Σ of the feature vectors belonging to class W.

Let L r-dimensional feature vectors belonging to class W be Xv
(j) = (xwl(j), xv2(j), -xvr
(j) ) (j=1.2...L).

Further, if each element of μm and Σ1 is defined as m k1σ(1)kl, it is calculated using the following equation by the mean/covariance matrix calculation device 14 shown in FIG. 5, which will be described later.

L j=1 The method for creating standard patterns 1 to M is shown in FIG.

As shown in the figure, to create a standard pattern,
A voice database 15 is created for each standard pattern class.

Specifically, the method for creating it is to first have multiple subjects pronounce the phonemes belonging to each class, and then record them.

A label is attached to the audio signal obtained in this manner frame by frame in order to distinguish between consonants and noise. This labeling involves displaying the waveform and spectrum of the audio signal on, for example, a CRT, and labeling each frame while viewing it. The label corresponding to this voice database 15 is defined as a label database 16.

Then, a feature vector is calculated from the speech database 15. Thereafter, referring to the label database 16, if the feature vector of the frame belongs to the class for which the standard pattern is to be created, the mean/covariance matrix calculation device 14 uses it to calculate the mean/covariance.

Therefore, the device of this embodiment creates a noise standard pattern using the feature vector estimated to be noise, and determines speech/noise based on this noise standard pattern and speech standard pattern. Even in large environment, a good detection rate can be obtained regardless of the input level.

Next, another embodiment of the present invention will be described.

ji! FIG. 6 is a diagram showing the configuration of the sound presence detection device according to this embodiment.

Feature vector calculation bag shown in the same figure! fl has the same configuration as shown in FIG.

The feature vector conversion device 17 is configured as shown in FIG.

The buffer 18 shown in the figure has a feature vector (X (n)
= (x 1(n), x2(n)...xp(n)
)) are accumulated in the buffer in the order in which the feature vectors are input into the buffer, in order to preserve the ordering relationship of the time when
The feature vectors are accumulated from the head of the buffer to the tail. That is, the temporally newest feature vector is stored at the head of the buffer, and the oldest feature vector is stored at the tail. The configuration of this buffer 18 is shown in FIG. As shown in the figure, the buffer 18 stores feature vectors of frames estimated to be noise by the threshold comparison device 19 (shown in FIG. 7). Therefore,
The feature vectors in the buffer 18 are not necessarily continuous in time.

Here, the set of feature vectors for N frames from the Sth frame of the head of the buffer 18 toward the tail is Ω(
n) and is expressed as follows.

Ω (n) −fXLn(S), XLn(S+1),
-XLn(S+N-1) )XLn(j) -(x
Lnl(j), x Ln2(j), -x Lnp
(j)) Also, the set Ω1 of the i-th element of XLn(・)
(n) is expressed by the following formula.

Ω1(n) = fxLni(S), xLni(S+1
).

---x Lnl(S+N-1)1 The threshold comparison device 19 compares the power of the feature vector X (n) with the threshold T (n) to estimate speech and noise. If the power component of X (n) is xi (n), if xi (n) ≧ T (n), then the sound is
If (n) < T (n), it is considered noise.

The threshold generation device 20 generates feature vector elements 2 for N frames from the input frame S frame past (S frame from the head of the buffer) toward the tail of the buffer from among the feature vectors accumulated in the buffer 18.
Take out the set ΩI(n) of 1 0 (power) and Ω1
Calculate the mean value and standard deviation of (n).

First, in the buffer 18, element Ml of the feature vector for N frames from the Sth frame of the head toward the tail.
O, and convert it into Ω1(n) −fxLnl(S
), xLnl(S+1).

---x　Lnl(S+N-1)) shall be.

Next, the average value ml and standard deviation σ1 are calculated for each element of the feature vector using the following equation.

N+5-1 m 1(n)-1/N, Σ x Lnl(j)N
+S-1 σ1 2-1/N-Σ (x Lnl(j)-m 1(n)) 2The average value m1(n) and standard deviation σ1(n) can also be written as below. .

ml(n)-ΣX 1N)/N σ12-Σ(x 1(j)-m 1(n)) 2
/Nj satisfies the following conditions, and the range of Σ is N frames that satisfy the following conditions from the larger j.

c x 1(j)eΩ(n)' ) &(j<n-
8) Here, Ω(n) is a set of feature vectors determined to be noise by the threshold comparison device 19.

The threshold value T (n) is calculated by the threshold generation device 20 as follows, for example.

T(n)=aXml+βxσ1 Here, α and β are arbitrary numbers.

However, until the feature vectors for N+S frames are accumulated in the buffer 18, T (n) is assumed to take a predetermined threshold value TO.

The conversion device 21 uses Ω(n) to convert the feature vector χ(n
) from which the effects of background noise and input signal level fluctuations are removed and set as Y(n). The transformation vector is an r (≦p) dimensional vector.

In order to calculate 'l(n), first, the average value and variance of each element of Ω(n) are calculated using the following equation.

N+5-1 m 1(n)-1/N, Σ x Lni(j)
N+S-1 cyi 2-17N' Σ (x Lni(j)-
m 1(n)) 2Here, Yr(n) is defined.

Vr(n) −(y l (n), y 2 (n), ・
..., y r(n)) Each element is defined by the following formula.

y 1(n) = (x 1(n) - m 1(n))
/ cr 1(n) Alternatively, the following formula may be used.

An example of calculating the y I(n) -x I(n) -m 1(n) transformation vector 'l (n) is shown below.

Example 1 Y (n) −V r(n) V(n) is the average vector M(n) of Ω(n) and X (
n) normalized by the variance of Ω(n). Here, ii, 2. ..., r, and r≦p.

Example 2 The transformation vector may be defined as the norm of 'I r(n).

Y (n) = II 'I r(n) II Here, 11·11 represents the norm of the vector.

Example 3 You may take the norm for each set of elements. For example, Y L(n) −(y 1(n), =...-, y r
l(n) )'l 2(n)-(y rl+1(n),
−・= , y r2(n) )”178(n) −(y
r2+1(n), --, y r(n))1 < rl
<r2<r, and using the norm of each vector, it is defined by the following equation.

Y(n)-(If'Yl(n)If, If Y
2(n)If, If'/3(n)If) output switch 22 converts the transformation vector 'l(n) estimated to be noise by the threshold comparison device 19 to the noise standard pattern creation device! Transfer to 7.

The noise standard pattern creation device 7 creates a noise standard pattern, and is configured similarly to that shown in FIG. However, here, the buffer 8 (see FIG. 3) stores the transformation vector Y (n).

The determination device 10 is also configured similarly to that shown in FIGS. 1 and 4. However, here, the voiced section is determined based on the transformation vector 'I (n) output from the feature vector transformation device 17. That is, matching device 1
1 is a conversion vector Y(n
), and if it matches the voice standard pattern, it is determined to be voice, and if it matches the noise standard pattern, it is determined to be noise.

Specifically, first, each standard pattern (μm, Σ
1) Measure the distance to (1-1, 2...N+1).

Di r)-('/-μm)1Σ1-'l'-μl)
+In Σ1 Y assumes that X belongs to the class w1 where i =tln Di ('/). If Wl belongs to audio, the frame is determined to be audio, and wl
If the frame belongs to noise, the frame is determined to be noise.

Here, the voice standard patterns stored in the voice standard pattern storage device 12 can be defined as follows.

The standard pattern is matching! 11, the black transformation vector belonging to class W has an average value vector μ and a covariance matrix Σ.

Let L r-dimensional transformation vectors belonging to class W be Yv
(j) −(yvlN), yv2(j), −yvr(
j) ) (j-1, 2...L).

Also, if each element of μm and Σ1 is m”k, σ(1) k, then the mean/covariance matrix calculation device!1
4 using the following formula.

m(1)k-1Σyvk(D L j-1 σ(I)kl-1Σ(xvkU) -m kL""
(xwl(j) −m I ) Therefore,
According to the device of this embodiment, the feature vector conversion device 17 removes the influence of noise from the voice feature vector and adaptively creates a standard noise pattern, so the S/N
A good detection rate was obtained even in an environment with large background noise where the ratio was about 20 dB to 14 dB. Furthermore, a detection rate independent of input level was obtained.

[Effects of the Invention] As explained above, according to the present invention, a noise standard pattern is created using only features/conversion vectors estimated to be noise, and speech/noise is determined based on this noise standard pattern. Therefore, even in environments with large background noise, a good detection rate can be obtained regardless of the input level.

[Brief explanation of drawings]

FIG. 1 is a diagram showing the configuration of a sound detection device according to an embodiment of the present invention, FIG. 2 is a diagram showing the configuration of the noise interval estimation device shown in FIG. 1, and FIG. 3 is a diagram showing the configuration of the noise interval estimation device shown in FIG. FIG. 4 is a diagram showing the configuration of the noise standard pattern creation device, FIG. 4 is a diagram showing the configuration of the determination device shown in FIG. 1, FIG. 5 is a diagram for explaining the voice standard pattern creation method, and FIG. 7 is a diagram showing the configuration of the feature vector conversion device shown in FIG. 6, and FIG. 8 is a diagram showing the configuration of the buffer shown in FIG. 7. 9 and 10 are diagrams showing the configuration of a conventional sound detection device. DESCRIPTION OF SYMBOLS 1... Feature vector calculation device, 2... Noise interval estimation device, 7... Noise standard pattern creation device, 10...
Judgment device. Applicant: Toshiba Corporation

Claims (2)

[Claims] (1) Feature vector calculation means that divides the input signal into frame units and calculates the feature vector of the input signal for each unit, and estimates the sound interval of this feature vector based on the feature vector calculated by the feature vector calculation means. a noise standard pattern creation means for creating a noise standard pattern based on the feature vector of a frame that is estimated to be not a sound period by the sound interval estimation means; It is determined whether the input signal for each frame unit is speech or noise from the feature vector calculated by the feature vector calculation means using the noise standard pattern created in advance and the speech standard pattern created in advance. 1. A sound detection device comprising: determination means. (2) A feature vector calculation means for calculating a feature vector of the input signal on a frame-by-frame basis from the input signal; and converting the feature vector calculated by the feature vector calculation means into a transformation vector from which fluctuations in the input signal are removed. A feature vector conversion means, a sound interval estimating means for estimating a sound interval for each frame based on the feature vector, and a noise standard pattern based on the converted vector estimated to be not a sound interval by the sound interval estimation means. A noise standard pattern creation means for creating a noise standard pattern creation means, a noise standard pattern created by the noise standard pattern creation means, and a pre-created audio standard pattern are used to determine whether the input signal for each frame is audio or noise. 1. A sound detection device characterized by comprising: determination means for determining whether a sound is present.