JPS63259694A - Sound source normalization - Google Patents

Abstract

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

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a sound source normalization method used, for example, in recognition processing of a speech recognition device targeted at unspecified speakers.

[Summary of the invention]

This invention is applicable to, for example, a speech recognition device for unspecified speakers! ! In the sound source normalization method used for identification processing, the average power value is subtracted from the output value of each channel obtained by frequency analysis, and the spectral envelope is shifted in parallel, so that the output value of the approximately central channel is approximately 0. The slope of the approximate straight line of the spectral envelope is calculated simply by addition and subtraction, and the sound source normalization process can be performed easily and efficiently using only the calculated slope of the approximate straight line. It is.

[Conventional technology]

Conventional speech recognition devices targeted at unspecified speakers are designed to be able to perform sufficient recognition processing on the voices of unspecified speakers other than those who have registered standard patterns.

In these speech recognition devices, in order to improve the recognition rate, it is necessary to normalize the overall tendency and dispersion of the frequency spectrum due to individual differences among speakers by some method. As a general normalization method, for example, a method is known in which the envelope of the frequency R spectrum is estimated by a linear function and normalized using the least squares method or the like.

For example, FIG. 3 shows an example of a spectral envelope based on data for one frame in time series obtained by frequency analysis of input audio. In Figure 3, the horizontal axis is the frequency (
In FIG. 3, the broken line 32 indicates the envelope, and each point on the broken line 32 indicates each output value from the frequency analysis section.

When performing normalization processing on such a spectral envelope, first, a straight line is estimated for which the sum of squares of errors is minimum for each output value on the broken line 32. In other words, if the number of channels in the frequency analysis section is n and the output value from each channel is xi{i=1...n), then the least squares approximation straight line indicated by 31 in FIG. 3 is y =a i +b, {i = 1---n)
−=(11, and at this time, the sum of squares f (xl) of the errors at each point on the straight line 31 corresponding to the output value x1 is 4° The slope that minimizes the sum of squares f (xl) of this error A least squares approximation straight line 31 is determined by a and the intercept S.

When actually calculating the slope a and the intercept, the relationship holds true, so the slope a and the intercept are i=l
i=1. Each output value x8 is normalized based on the obtained least squares approximation straight line 31. That is, normalized Xi = X
It is calculated by l - (a i + b) = (5), and this subtraction process flattens the overall tendency and dispersion of the spectrum due to individual differences among speakers.

[Problem that the invention seeks to solve]

However, in the conventional normalization method using the least squares method described above, it is necessary to find the slope a and the intercept of the least squares approximation straight line, and in addition, multiple multiplications are required in the process of these calculations. must be carried out. For this reason, there is a problem that the number of software steps for calculation processing increases and the processing time becomes longer.

Therefore, an object of the present invention is to provide a sound source normalization method that can easily and efficiently normalize the tendency of a frequency spectrum based on an approximate straight line.

[Means for solving problems]

In this invention, the sum of powers P1 from channel 1 to channels up to the center and n from channels above the center are calculated.
The sum of the power up to the channel P2 and the sum of the entire power (
a step of calculating the average value m of power from Pl+P;
A step of subtracting the average power value m from the output value of each channel to obtain Ym, and a slope of the approximate straight line such that the power of the average value has a sum of powers equal to the sum of the powers of the subtracted output value Xi. a and the normalized value Xl for n channels as Xi −xi−a {i −(n+ 1)/2)
The normalization process is performed by the step of obtaining −=(9).

[Effect]

The average power value m is calculated from the output value X of each channel obtained by frequency analysis, and the average power value m is subtracted from the output value X of each channel to calculate the initial normalized value Y4. Through this subtraction process, the spectral envelope of one frame is translated in parallel so that the output value becomes approximately 0 in the channel located approximately at the center, and the approximate straight line of the envelope drawn by the initial normalized value Xi is y = a (1-(n + 1) / 2) (
n: number of channels, i: channel number 1≦i≦n)
......(8), and the slope a of the approximate straight line is the sum P+ of the power from channel 1 to the approximately central channel, and the phase P2 of the power from approximately the central channel to channel n. It is calculated using only addition and subtraction. The final normalization process is performed using the slope a of the obtained approximate straight line, and the normalized value Xi is xl = xl - a (1-(n+1)/2) = xl
−m−a {i −(n+1) /2) {i=1.・・・
... n) ... Calculated by (9).

〔Example〕

a. Configuration of speech recognition device and its processing flow An embodiment of the present invention will be described below with reference to the drawings. FIG. 2 conceptually shows an example of the flow of processing carried out in a speech recognition device, and the present invention relates to the sound source normalization processing indicated by 15 in FIG.

In FIG. 2, the part indicated by 11 represents the human-powered system. In the input system 11, preprocessing required for speech recognition processing is performed. For example, input audio is supplied to an amplifier via a microphone, and the input audio signal is amplified and supplied to a low-pass filter. The input audio signal is limited to the band required for tuft processing in the low-pass filter.

Then, the input audio is converted from analog to digital at a predetermined sampling frequency. A digital audio signal from an input system 11 is supplied to an analysis system 12 .

The analysis system 12 includes, for example, a digital band-pass filter bank made up of n band-pass filters, and performs frequency analysis on the input audio signal. For example, the center frequencies of each passband of a digital bandpass filter bank are distributed at equal intervals on the logarithmic axis, and the n-channel frequency obtained by supplying an input audio signal to this digital bandpass filter bank is Each of the outputs is squared and then averaged to form a power spectrum.

Therefore, the audio signal is expressed by the magnitude of the power spectrum of n channels that are equally spaced on the logarithmic axis. Then, a data string indicating the power spectrum of n channels is outputted as one frame every unit time (frame period). That is, the audio signal is cut out as a parameter expressed by an n-dimensional vector every frame period, and is supplied to the recognition processing system 13.

The recognition processing system 13 includes, for example, a special@quantity extractor, a speech interval determiner, a sound source normalizer, and a reject determiner.

It is composed of a NAT processing section, a matching judgment section, etc. Note that the portion surrounded by the broken line in FIG. 2, which is performed by the recognition processing system 13, is processed frame by frame.

First, a parameter conversion process 14 is performed on the n-channel power spectrum of each frame from the analysis system 12, and, for example, the power spectrum is logarithmically transformed into a logarithmic power spectrum.

Then, sound source normalization processing 15. Voice section determination processing 16,
Each process of the feature amount extraction process 17 is performed.

In the sound source normalizer, an approximate straight line to the spectrum envelope is estimated for each frame, and the sound source normalization process 15 is performed using this approximate straight line. In the sound source normalization method of the present invention, the slope of the approximate straight line is calculated only by addition and subtraction, and the sound source normalization process 15 is performed without calculating the intercept component of the approximate straight line.

Further, a feature extraction process 17 is performed in the feature extractor. For example, the number of zero crossings of the input audio signal is counted in frame units, and the count value is determined, and the power of the input audio signal in each frame, that is, 2
The sum of the products is calculated.

At the same time, the phoneme characteristics of each frame, that is, the characteristics of the shape of the spectral envelope, etc. are detected.

Data indicating the feature amount obtained through these processes is newly added to each frame as a parameter.

Further, in the speech segment determination device, a complex speech segment determination process 16 is performed based on the count value of the number of zero crossings, the power of each frame, and the sound source normalization information, and for example, silence, voiceless sound, and voiced sound are determined. A voice interval is determined.

At this time, a reject process 21 is performed in the reject determiner in order to distinguish between ambient noise and the input voice. For example, the power level of each frame is compared with a predetermined threshold, and when it is found to be greater than the predetermined threshold, it is assumed that speech has been input and the speech section determination processing 16 and special f&I extraction processing 17 are performed, When it is smaller than a predetermined threshold, it is judged as ambient noise etc. and is rejected.
It is considered an invalid input.

Only the frames corresponding to the voice interval determined by the voice interval determination process 16 are considered valid to form a characteristic pattern, and NA T (Norma) is applied to this characteristic pattern.
lization Along Trajecto
ry) Process 18 is performed. That is, normalization processing is performed along a time-series trajectory in the feature vector (corresponding to the number of parameters, and is represented by N parameters, an N-dimensional vector) space, and the feature pattern is transformed in the time axis direction. For example, the inter-frame distance between adjacent frames that make up the feature pattern is calculated, and the sum of the inter-frame distances is then calculated to calculate the trajectory length from the starting frame to the ending frame of the feature pattern. is required. Then, the trajectory length is divided into equal parts by a predetermined number of divisions required to extract the features of the feature pattern, and only frames that are close to each division point are extracted to generate the speaker's voice. The time axis is normalized so that it is not affected by speed fluctuations.

The feature pattern that has been subjected to the NAT processing 18 is subjected to 2-bit conversion processing 19, and, for example, each data in each frame constituting the feature pattern is converted to 2 bits, thereby compressing the amount of data.

Matching processing 20 is performed between the standard pattern registered in advance and the input voice characteristic pattern, for example, pattern matching is performed between all the standard patterns selected as comparison targets.

For example, inter-frame distances are determined between frames that make up the characteristic pattern and frames that make up the standard pattern, and the sum of the distances is taken as the matching distance.

At this time, reject processing 21 is performed in the reject determination device.
For example, the matching distance calculated between each standard pattern and 9 is compared with a predetermined threshold,
Those that are greater than a predetermined threshold are rejected as not applicable. Then, a determination process 22 is performed to determine the minimum matching distance that is smaller than a predetermined threshold, and the word corresponding to the standard pattern with the minimum matching distance is determined as a recognition result.

b. Description of sound source normalization processing The above-mentioned sound source normalization processing will be explained with reference to FIGS. 1A to 1C. In each of FIGS. 1A to 1C, the horizontal axis represents the frequency (channel), and the vertical axis represents the level of the power spectrum.

For example, FIG. 1A shows an example of a spectral envelope based on one frame of time-series data obtained by frequency analysis of input audio. In FIG. 1A, the solid line indicated by 1 indicates the spectral envelope. Each point on the solid line indicates the output value of each channel (for example, where n is the number of channels, n=16) obtained by frequency analysis. In addition, the solid line indicated by la in FIG. 1A is an approximate straight line for the spectrum envelope 1.

The average value m of the power at each point on the spectrum envelope 1 in FIG. 1A is calculated as follows, where X is the output value of each channel. Using this average value m, the process shown in (7) 2 below is performed. That is, the average power value m is subtracted from the output value Xi of each channel to calculate the initial normalized value Y.

Xi =x, -m {i-1+...n)・
(7) By the process shown in the above equation (7), the spectral envelope 1 and the approximate straight line 1a are translated in parallel as shown in FIG. 1A, and are made into the solid lines 2 and 2a. At this time, the approximate straight line 2a intersects the X axis at the approximately central position of the channel, so y=a {
i − (n+1)/2) (n: number of channels, i: channel number 1≦i: 5n
) ...(8) can be assumed. Therefore, the final normalized value Xi
is Xt −Xt a {i−(n+1)/2)Xi
−m−a {i (n+1) /2) {i=1.
. . . n) . . . It can be calculated by (9).

For example, assume that the output (! Assuming that the area of region 4 is 81, the area S1 is calculated by (in addition, the [ ] Nygauss symbol).Also, by the approximate straight line indicated by the solid line 5 in Figure 1C, which satisfies the above-mentioned equation (8), Assuming that the area of the hatched region 6 formed is 82, the area S2 is calculated by Oυ.It can be assumed that these areas s1 and 82 are equal, and the approximate straight line 5 The slope a is calculated as follows.

The denominator on the right side of the above equation (b) is a constant because the number of channels n is fixed. Therefore, the slope a of the approximate straight line is
Calculated by multiplying the difference between the sum PL of the output values in the first half from channel 1 to the channel located approximately in the center and the sum P2 of the output values in the second half from the channel located approximately in the center to channel n. be done.

In other words, the slope a of the approximate straight line can be found by only adding and subtracting without performing multiplication on the output of each channel.
The final normalized value Xl is calculated by the following equation (9)'.

XI −XI −a {i −(n+1)/2)x
{i − (n+1)/2) ・Old... (9) In actual calculations, the number of channels in the analysis system 12 is an even number (n
= 2m), but the calculation process is slightly different depending on whether it is an odd number (n-2m+1), and each case will be described below.

i) The number of channels n is an even number (n=2m.

In the case of m ” 1 * 2 +...) (as described above (
9) The denominator of the second term on the right side of the equation is i=rigid+l
i=1=m".

x {i −(2m+1) /2) ,'−2n+” xt =h” Xi 2k i
+ (211+1) k,', S+ Xz =
S+X=St l +S3.

In addition, S+ = 2n+”.

s, = (2o+I)k.

ii) The number of channels n is an odd number, (n=2m+1,
m=1.2.・Old...), the denominator of the second term on the right side of equation C9 above) is -m(m+1
).

x {i − (m+1) ) , ゛, II (1m+1)XI = m(m+1)
xa −k ” i +k ” (m+1
), ', S+'X1-5+'Xl5z'i+ss'.

In addition, S 1 ′ = rn (m + l)33
”=(m+1)k'.

Note that in both cases where the number of channels n is an even number (n = 2 m) and an odd number (n = 2 m + 1), it is taken as the form Sl (or S, ')
is multiplied by a constant, but in the recognition process, since it is a relative comparison, it does not affect the recognition rate in any way, and s,
(or S, ′)Xxz is a constant times T1 and S
3 (or S, ′〉) to successively s2 (or 82
′) is calculated by subtracting the amount.

Also, if the number of channels is odd (n = 2 m + 1
), the sum P of outputs is calculated with channels 1 to channel (m+1) as the first half, and the sum Pt of outputs is calculated as the second half from channel (m+2) to channel n for normalization. However, when the number of channels is odd, the sum of the outputs P+ and Pg may be obtained by using the output value of the (m+1) channel located in the center for both calculations. The output sums Pr and Pg may be determined while ignoring the output values of the (m+1) located channels.

〔Effect of the invention〕

In this invention, the average power value m is determined from the output value Xi of each channel obtained by frequency analysis, and the output value X! of each channel is calculated. An initial normalization process is performed by subtracting the average value m of power from . With this initial normalization, the spectral envelope of one frame is translated in parallel so that the output value in the channel located approximately at the center becomes approximately O, and the slope a of the approximate straight line changes from channel 1 to approximately the center. It is calculated by an equation consisting only of addition and subtraction using the sum P1 of the powers up to channels equal to n and the sum P2 of powers from the approximately central channel to n channels.

The final normalization process is performed using the slope a of the obtained close-up straight line.

Therefore, according to the present invention, there is no need to calculate the intercept, which was necessary for the normalization process using the conventional least squares method.
The tendency of the frequency spectrum can be easily and efficiently normalized using only the slope a.

Further, according to the present invention, the slope a of the approximate straight line used in the normalization process can be calculated only by addition and subtraction,
Furthermore, it is possible to efficiently normalize the tendency of the frequency spectrum.

For reference, a comparison of the amount of computation between the conventional normalization process using the least squares method and the normalization process according to the present invention for one frame will be described.

When performing normalization processing using the least squares method, the above (
4) As shown in the formula, IXxl, {i==11・−・・−n)・
−・−・−(J4J multiplication is executed n times, and ΣixΣX! ... River α is executed once to determine the slope a and the intercept S. Then, the normalized value is calculated. ax+ at stage
. . . It is necessary to perform the multiplication αe n times.

On the other hand, according to the normalization process of the present invention, since normalization is performed only with the slope a calculated only by addition and subtraction, multiplication corresponding to the above α-y formula and 051 formula is unnecessary, and the above αω formula By performing the multiplication ax {i − (n+1)/2) corresponding to only once, the normalized value can be obtained and is processed very efficiently.

[Brief explanation of drawings]

1A to 1C are route maps used to explain an embodiment of the present invention, FIG. 2 is a conceptual diagram as an example used to explain a speech recognition device, and FIG. 3 is a diagram used to explain a conventional sound source normalization method. This is the route map to be used. Agent Patent Attorney Masato Sugiura Figure 1

Claims (1)

[Claims] The sum P_1 of the power from channel 1 to the center channel, the sum P_2 of the power from channels above the center to n channels, and the sum of the total power (P_1+P_
2), a step of calculating the average power value m from the output value of each channel, a step of subtracting the above average power value m from the output value of each channel to obtain @x@_i, and an output value @ from which the power of the above average value is subtracted. A step of calculating the slope a of an approximate straight line having a sum of power equal to the sum of powers of x@_i, and a normalized value for n channels ■_i according to the following equation (9).
A sound source normalization method characterized by comprising a step of determining . ■_i=@x@_i-a{i-(n+1)/2}...
...(9)