JPH0348900A - Sound detecting device - Google Patents

Abstract

PURPOSE:To improve the detection accuracy even in the case a background noise is large by accumulating a feature parameter of a frame unit voice in a buffer, and forming a necessary conversion parameter from the feature parameter and deciding it. CONSTITUTION:A feature parameter of a frame unit sound signal is calculated by a feature parameter computer 1 and accumulated in a buffer of a feature parameter converter 2. Subsequently, a distance vector of an assembly of feature parameters of the past N frames based on the feature parameter before S frames from an input frame containing the feature parameter of a noise period which is accumulated, as a reference and the feature parameter of the input frame is calculated, and the feature parameter of the input frame is converted to the conversion parameter which is not influenced by a noise. By using this conversion parameter, a sound is decided by a sound deciding device 3 and even in the case a background noise is large, the voiced sound detection accuracy is enhanced.

Description

[Detailed Description of the Invention] [Object of the Invention] (Industrial Application Field) The present invention provides an ATM (Asynchronous T
transfer mode) communication, DSI (Digi
tal 5peech Interpretation
), it is applied to the fields of packet communication and voice recognition, and relates to a sound detection device that accurately detects sound intervals in a voice signal.

(Conventional technology) FIG. 10 shows a configuration of a conventional sound detection device.

Feature parameters such as power, number of zero crossings, autocorrelation function, spectrum, etc. are calculated by the feature parameter calculator 101 on a frame-by-frame basis from the audio signal input to the input terminal 100.

The calculated feature parameters are output to the matching device 102, and compared with a preset voice standard pattern 103 and a noise standard pattern 104, and the respective distances are calculated.

If the distance between the feature parameter and the sound standard pattern 103 is smaller than the distance between the feature parameter and the noise pattern 104, it is determined that the input frame belongs to the sound pattern, and if the opposite is true, it is determined that the input frame belongs to the noise, and the determination result is Output terminal 10
Output from 5.

(Problems to be Solved by the Invention) However, unlike vowels, the power of consonants is often lower than the power of background noise even if they are voiced.

For this reason, in an environment with large background noise, the characteristics of the background noise will appear significantly in the characteristic parameters of the consonant section.

According to the conventional sound detection device described above, the feature parameters affected by background noise are used as they are for determination.
When the background noise was large, there were many consonant detection errors.

This has caused a deterioration in sound quality in the field of communications, and has caused a decline in recognition rates in the field of speech recognition.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a voice detection device that can improve the accuracy of detecting presence of voice even when background noise is large.

[Structure of the Invention] (Means for Solving the Problems) In order to solve the above problems, the first invention provides means for calculating voice characteristic parameters on a frame-by-frame basis from an input voice signal; A buffer for sequentially accumulating all feature parameters for each frame or feature parameters for a noise interval, and a buffer for sequentially accumulating feature parameters for each frame or for a noise interval, and a buffer for sequentially accumulating feature parameters for the past N frames based on the feature parameters for S frames before the input frame from among the accumulated feature parameters. means for generating transformation parameters for the feature parameters of the input frame by extracting a set and calculating a distance vector or norm of the vector with the feature parameters of the input frame; and comparing the transformation parameters with a preset standard pattern. and means for determining a sound interval in the audio signal.

The second invention also provides a means for calculating audio feature parameters in frame units from an input audio signal, a buffer for sequentially accumulating the calculated feature parameters in frame units, and regression from the accumulated feature parameters. means for calculating a coefficient to generate a conversion parameter of a characteristic parameter of an input frame; and means for comparing the conversion parameter with a preset standard pattern to determine a sound section in the audio signal. Features.

(Operation) In the above configuration, in the first invention, the distance vector or Calculate the norm of the vector and use these as transformation parameters. Then, by comparing this conversion parameter with a standard pattern, the presence of a voice is determined while avoiding the influence of noise.

Furthermore, in the second invention, a regression coefficient is determined from the accumulated feature parameters, and a conversion parameter is determined from this regression coefficient. Then, by comparing this conversion parameter with the standard pattern, the sound presence is determined while avoiding the influence of noise.

(Embodiment) FIG. 1 is a block diagram showing a schematic configuration of a sound detection device according to the present invention, which includes a feature parameter calculator 1, a feature parameter converter 2, and a sound presence determiner 3. It consists of

In each of the following embodiments, the audio signal is analyzed frame by frame to determine the presence or absence of audio.

For example, if an audio signal is sampled at 8KHz and
Each sample is combined into one frame.

However, the frame length does not always have to be a constant length.

The feature parameter calculator 1 calculates Durb for each frame.
Calculate linear prediction coefficients using the in method or the like. here,
A PARCOR coefficient, LPC cepstrum, mel cepstrum, etc. may be calculated from the linear prediction coefficients and used as feature parameters. Power, autocorrelation function, number of zero crossings, etc. may also be calculated.

The frame for which it is currently being determined whether there is a sound or no sound is hereinafter referred to as an input frame. In addition, the feature parameters of the input frame obtained by the feature parameter calculator 1 are set to X(n
). n is the sequential number of the frame. The feature parameter is a p-dimensional vector, as shown in (1) below.
It is written as the formula.

X(n) = (xt (n), x2(n)s=,
XP (n)) (1) The feature parameter converter 2 converts feature parameters in order to emphasize the difference between speech and noise.

The feature parameters converted here are hereinafter referred to as conversion parameters, and can be expressed by the following equation (2).

The change parameter is an r (≦P)-dimensional vector.

Y(n) = (yt (n), y2(n), =-, y
r (n))...(2) FIG. 2 is a block diagram showing the configuration of the feature parameter converter 2, which is a characteristic part of an embodiment of the present invention, and this embodiment avoids the influence of background noise. In order to
After determining the conversion parameter Y(n), the conversion parameter Y(n) is compared with a standard pattern to determine a sound interval.

The feature parameter converter 2 of this embodiment includes a power measuring device 4, a noise determiner 5, a switch (SW) 6, and a buffer 7.
and a distance vector calculator 8.

The power measuring device 4 measures the average power P on a frame-by-frame basis using the following equation (3).

The samples of the audio signal in the frame are a(1)(1-0,
1, N-1), and the number of samples in one frame is N, then the average power P = "? a(1)2/ N -
(3) The noise determiner 5 compares the input signal with the power P measured with a power measuring device having a given threshold value T in order to detect a section that is definitely noise from the input signal.

If P≧T, it is determined that it is not noise and outputs “0” to SW6.

If not, it is determined to be noise and "1" is output to SW6.

If the output of the noise determiner is "1", SW6 causes the buffer 42 to store the characteristic parameters of that frame.

In the buffer 7, the feature parameters are accumulated from the head of the buffer to the tail in the order in which they were input to the buffer, in order to preserve the order of the time in which the feature parameters are accumulated in the buffer 7.

That is, the newest feature parameter (the feature parameter of the frame to be currently determined) is stored in the head of the buffer, and the oldest feature parameter is stored in the tail. FIG. 3 shows an example of the configuration of the buffer 7.

The distance vector calculator 8 extracts a feature parameter set Ω for N frames from S frames before the input frame (S frames from the head of the buffer) to the tail of the buffer from among the feature parameters accumulated in the buffer 7. , Ω and the feature parameter X(n) of the input frame.

Note that the S frame and the N frame can be any number of frames, but preferably several frames to about 20 frames.

First, in the buffer 7, a set Ω of feature parameters for N frames past, counting from the frame S frames before the input frame, is expressed as Ω: XL (0), X
Let L (1), ..., X, (N-1,) +.

Let the noise distance vector between X(n) and Ω be Y(n).

Y(n) is the difference between the average value M of Ω and X(n) normalized by the standard deviation of Ω, and is expressed by the following equation (4).

FIG. 4 shows the relationship between X(n), Ω, and Y(n).

Y(n) = (y+ (n), y2(n), -, 3'
q (n))...(4) M" (mt 1m2* ・"1mq)...
(5), then Y+ (n) −(x+ (n), mt ) /at
-(6)...(8) Here, 1=I-, 2,-, q, when r > 1, q
= r and the transformation parameter is Y (n).

When r=1, q can take any size in the range 1≦q≦p. In particular, when r=1 and p>l, 11.
11 is the norm of the vector, then the transformation parameter is I
IY(11) +1. ! 1. do.

The sound determination unit 3 determines a sound interval based on the conversion parameters obtained from the feature parameter converter 2. As shown in FIG. 5, the utterance determiner 3 is composed of a matching device 9 and M standard patterns 10.

The matching device 9 measures the distance between the standard pattern and the conversion parameter, and determines that if the standard pattern is matched with the standard pattern belonging to audio, it is audio, and if not, it is determined to be silence.

First, each standard pattern ω, (i=1°...
・, M).

DI (Y) = (Y−μ()′ Σ+ −′ (Y−μ,)
+101 Σ 1
...(9) Y is i = min D + (Y)
...(10) Suppose that Y belongs to ω1. If ω1 belongs to voice, then the frame is voiced, ω,
If the frame belongs to noise, the frame is determined to be noise.

Standard pattern 10 can be defined as follows.

The standard pattern 10 is based on the configuration of the matching device 9, and i, which indicates the class of the standard pattern, is omitted for the sake of simplicity.

Let L r-dimensional transformation parameters belonging to class ω be Y,
(j)-(y-t(j), y, 2(j),...
y,, (j))...(11) Let j-1, 2,..., L.

Further, assuming that the respective elements of μ and Σ are μ1 and Σ1, the calculation is performed by the mean/covariance matrix calculator 8 in FIG. 5 using the following equation.

((yWt(j) - μL) (13) The standard pattern (standard pattern M) is created as follows.

To create standard patterns, audio data is created for each class of each standard pattern.

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 on a frame-by-frame basis to distinguish between consonants and noise. Labeling is done by CR of the waveform and spectrum of the audio signal.
T, and label each frame while looking at it.

From the labeled speech database 11, conversion parameters are calculated as shown in FIG.

At this time, referring to the label database 12, if the conversion parameter of the frame belonging to that class is the average
A covariance matrix calculator 13 calculates the average and covariance.

The standard pattern created in this manner is compared with the conversion parameters determined as described above, and a sound section that avoids the influence of noise is determined.

In the above embodiment, the power measuring device 4, the noise determining device 5, and the switch 6 are used in order to determine the noise in advance and store it as the characteristic parameter of this section. However, when the noise is not determined, The feature parameters of all sections may be stored in the buffer 7. In this case, the power measuring device 4, noise determining device 5, and switch 6 are unnecessary.

In addition, a power measuring device 4 and a noise determining device 5 were used to determine noise, but similar to the presence determining device shown in FIG.
The noise section may be determined using a standard pattern of noise and a matching device.

FIG. 7 is a block diagram showing the configuration of the feature parameter converter 2, which is a characteristic part of another embodiment of the present invention. The data is accumulated, an avoidance coefficient is calculated from the accumulated data, and a conversion parameter is determined from this regression coefficient to determine a sound interval.

The buffer 14 has the same configuration as the buffer 7, and sequentially accumulates feature parameters on a frame-by-frame basis.

The regression coefficient calculator 15 calculates regression coefficients and conversion parameters as follows.

Let the feature parameters of the input frame be X(n), and each element X+(k) of the p-dimensional feature parameter vector of the th frame in the buffer is given in the range of (n-(N-1))≦≦n. Then, each element x+(k) has (n-(N-
1)) ≦ to ≦n, the approximation function is f + (k)-a
to + a ILK” + a 12 to 20...+
Assume that a, , ni'... is expressed as (14). This is shown in FIG.

The difference between x+(k) and f+(k) at each point, that is, the residual e+ (k) = f+ (k) −Xi (k)
- If the sum of squares of (t5) is Jl, J becomes J, = Σ e l 2(k). Each regression coefficient aII is set so that this Jl is minimized.
) + ..., a, is determined.

The conversion parameter of the input frame is the following formula (17) % formula % (17) or a combination of a plurality of i's. For example, Y (n)
” (az+ a21+ ”'+ arl+ a
12+a 22. -, a r2) ...(
[1] becomes.

The conversion parameter Y(n) obtained in this manner is compared with a standard pattern by the sound presence determination unit 3 to determine the sound interval. It should be noted that the determination of the voiced section is performed by the same process as in the previous embodiment, so the explanation will be omitted.

The effects of each of the above embodiments will be explained based on specific measurement results.

Unlike vowels, the power of consonants is often lower than the background sound power. Therefore, in an environment with large background noise, the characteristics of the noise will appear significantly in the feature parameters even in consonant intervals.

In the conventional method, feature parameters affected by background noise are used as they are for determination, so when the background noise is large, consonant detection errors increase.

In each of the embodiments of the present invention, in order to emphasize the characteristics of noise and voice, a good detection rate was obtained even in an environment with large background noise, where the S/N ratio was between 20 dB and 14 dB.

Below, the detection results of word-initial consonants when changing the feature parameters and feature parameter conversion methods are shown.

If it is determined that a frame whose label attached to the audio data indicates a consonant is in one of the consonant classes, it is determined that the frame has been correctly detected.

The detection rate shown in FIG. 10 is the average value of the consonant detection rate and the noise detection rate. The consonant detection rate is defined by the following equation.

Consonant Detection Rate - Furthermore, if it is determined that the frame of noise data belongs to any of the noise classes, it is assumed that it has been correctly detected. This is the noise detection rate and is defined by the following equation.

Noise detection rate= In FIG. 10, the vertical axis is the detection rate. In addition, the horizontal axis indicates the type of feature parameter, where LPC is the LPC cepstrum, P is the average power within the frame, and P+LPC is the P
and LPC.

In addition, in the following, the LPC cepstrum analysis dimension is 12th,
The conversion parameter dimension is 4 when the feature parameter is LPG.
Next, when P+LPG, it was set to 5th order.

The feature parameter conversion method is shown by changing the plot.

C is a conventional method that does not perform feature parameter conversion.

n is the embodiment shown in FIG. 2, in which noise is determined.

2 is the embodiment shown in FIG. 2, in which noise determination is not performed.

r is another example shown in FIG. 7, in which feature parameters are converted into regression coefficients.

[Effects of the Invention] As explained above, according to the present invention, the influence of noise can be removed from the feature parameters by feature parameter conversion, so that it is possible to accurately determine a sound interval even in an environment with large background noise. I can do it.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing a schematic configuration of a sound detection device according to the present invention, FIG. 2 is a block diagram showing characteristic parts of an embodiment of the present invention, and FIG. 3 is a buffer used in the embodiment. Fig. 4 is an explanatory diagram of the conversion parameters of the same embodiment.
FIG. 5 is a block diagram showing an example of the configuration of a voice determination device, FIG. 6 is a block diagram showing an example of the configuration of a device for creating a standard pattern, and FIG. 7 is a block diagram showing a characteristic part of another embodiment of the present invention. The block diagram, FIG. 8 is a diagram showing the approximation function in the process of calculating the regression coefficient in the same embodiment, FIG. 9 is a characteristic diagram showing the relationship between the feature parameters and detection rate in each embodiment, and FIG. 10 is the conventional FIG. 2 is a block diagram showing a configuration example of a sound detection device. 1... Feature parameter calculator 2... Feature parameter converter 3... Sound determiner 4... Power determiner 5... Noise determiner 6... Switch 8... Noise vector calculation Device 9...Matching device 10...Standard pattern

Claims (2)

[Claims] (1) A means for calculating voice characteristic parameters in frame units from an input voice signal, a buffer for sequentially accumulating all the calculated characteristic parameters in frame units or characteristic parameters of noise intervals, Among the feature parameters, a set of feature parameters for past N frames is extracted based on the feature parameters S frames before the input frame, and the distance vector or norm of the vector from the feature parameters of the input frame is calculated. a means for generating a conversion parameter of the characteristic parameter; and a means for comparing the conversion parameter with a preset standard pattern to determine a sound section in the audio signal. Detection device. (2) means for calculating audio feature parameters in frame units from an input audio signal; a buffer for sequentially accumulating the calculated feature parameters in frame units; and calculating regression coefficients from the accumulated feature parameters; An apparatus characterized by comprising: means for generating a conversion parameter of a characteristic parameter of an input frame; and means for comparing the conversion parameter with a preset standard pattern to determine a sound interval in the audio signal. Sound detection device.