JPH035595B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition method using dynamic features of speech.

In speech recognition, methods using dynamic programming have conventionally been widely used for pattern matching using dynamic features of speech. According to this method, when comparing the input pattern and the standard pattern, the speaking speed is usually different, so in order to maximize the similarity between the patterns, the input pattern is first subjected to time stretching processing such as stretching or contracting. After such processing is performed, the degree of similarity between patterns is determined.

The present invention aims to further improve the above-mentioned recognition method using the dynamic features of speech. In particular, the present invention utilizes the dynamic features of speech by using a dynamic model in the feature vector, and further improves the recognition method using the dynamic features of speech. By performing optimal estimation using an evaluation function with
To provide a speech recognition method that can reduce the influence of fluctuations in the values of feature vectors directly obtained from input speech, increase the degree of separation between categories of each speech, and facilitate matching with standard patterns.
Next, the present invention will be explained in detail with reference to Examples.

Before the input audio information is compared with a standard pattern, the temporal changes in the input audio spectrum are expressed using dynamics (set on page 4) in order to dampen the fluctuations in the value of the feature vector. The optimal spectrum is estimated for each frame using an evaluation function called a Kalman filter, and pattern matching is performed using this estimated spectrum in place of the actually measured spectrum. The principle for obtaining the optimal estimate from input speech will be explained.

Note that, similarly to the conventional method, time expansion/contraction processing is performed in advance on the measured input audio in order to match it with the standard pattern.

Audio input applied to an audio input unit such as a microphone is applied to an analysis unit, where it is sampled in a predetermined interval and converted into a digital signal to form a characteristic pattern of the audio. Here, a pattern is a time series of feature vectors, and can generally be expressed as a time series in a multidimensional space. For example, the output of frequency analysis of an input audio waveform via a bandpass filter or the like can be expressed as a vector having output values in each frequency band as elements,
This becomes the feature vector described above. Therefore, if the frequency domain is divided into n bands (n=24 in this embodiment), the feature vector becomes n-dimensional, and the pattern becomes a time series of n-dimensional feature vectors X ₁ X ₂ . _. . In the following, bold fonts represent vectors.

The feature vector in the i frame in the pattern can be expressed as a time change with respect to the (i-1) frame by the following difference equation.

X _i =A _i-1 X _i-1 +U _i-1 (1) Here, the parameter A _i-1 is a matrix that describes the time change of X _i-1 , and is assumed to be a diagonal matrix for simplicity. Also
U _i is an excitation source and is expressed as a time variation of a smooth spectral outline, which will be described later.

For the true feature vector 〓 _i as described above, the vector 〓 i actually obtained from the measured values 〓 _i includes an error 〓 _i and is expressed as follows.

Y _i =X _i +W _i (2) The expressions in equations (1) and (2) above are merely examples and do not limit the field of application of the present invention.

In the process of speech recognition, the above-mentioned measured value Y _i is obtained, and an optimal estimated value X^ _i of X _i is determined from this measured value Y _i using an evaluation function having a predetermined algorithm. A Kalman filter is used as the above evaluation function.

Since we are currently using a Kalman filter, the evaluation function is a squared error, and the above optimal estimate X^ _i is a vector of estimated values that minimizes the squared error, and the least squares estimated value X^ _i is calculated according to the following algorithm. It is required.

X^ _i =X〓 _i +P _i W _i ^-1 (Y _i −X〓 _i ) (3) X〓 _i =A _i-1 X^ _i-1 + _i-1 (4) P _i = (M _i ^-1 +W _i ^-1 ) ^-1 (5) M _i =A _i-1 P _i-1 A _i-1 +U _i-1 (6) where W _i is the covariance matrix of W _i , and for simplicity
The average value of W _i is set to zero, and the variance of each element of W _i is constant regardless of time and is set to σ ² . Also, let W _i = σ ₂ I (I is the identity matrix), U _i is the covariance matrix of U _i , the average value of U _i (U _i = g _{i +1} − g _i ) is U _i , and P _i is the covariance matrix of estimation errors.

Figure 1 shows a dynamic model of feature vectors X _i , X _{i+1 , and X i+2 in frames i, i+1} , and _i+2 . The feature vectors are vectors consisting of output values of 24 bands, and each element value is Each frame is indicated by a white circle. From the figure, the excitation source U _i is determined from U _i = g _{i +1} − g _i by using the time change of the spectral outline g _i shown by the broken line, and the time series of the feature vector is ( 1) It can be expressed without contradiction according to Eq.

FIG. 2 shows the processing procedure for obtaining the feature vector X _i and the spectral outline g _i .

First, a sampled and digitized audio waveform is divided into predetermined time intervals (frames) (for example, 10 to 20 ms) and analyzed for each interval data. A weight (here, a weighting coefficient called a Hanning window) is applied to the data of each frame according to the time position of the data by windowing processing.

Thereafter, a power spectrum is calculated by digital Fourier transform (DFT), and logarithmization (LOG) is performed for each spectral component to obtain a logarithmic spectrum (LOG SPECTRUM). Divide this spectrum into bands (PASSBAND WINDOW)
By compressing it into a small number of spectral components (here, 24 channels), a feature vector X _i is obtained.

On the other hand, cepstrum coefficients (CEPSTRUM) are calculated from the logarithmic spectrum by inverse Fourier transform (IDET), and then digital Fourier transform (DFT) is performed using only the low-order coefficients (in this case, 5th order) (WINDOW). , obtain the envelope spectrum of the logarithmic spectrum. Divide this envelope spectrum into bands (PASS BAND
WINDOW) (here 24 channels), we obtain an approximate vector g _i with smooth characteristics.

_Time series of feature vectors of the current speech category α ₍ X〓 ₁ , _{X〓 2} ...

〓k _-1 ) is obtained in advance as a standard pattern. In this state, the time series of the parameter A〓 _i regarding this category α can be obtained from equation (1).

On the other hand, from the input speech, the time series of feature vectors (Y ₁ , Y ₂ ... Y _k ) and excitation sources (U ₁ , U ₂ ...
U _k-1 ) can be found, and (〓 ₁ 〓 ₂ …〓 _k-1 ) (A〓 ₁ A〓
₂ ...
A vector of estimated values (X^ _{1 X} ^ ₂ . _. .

_{Time series of feature vectors ( X ^ 1} is the measured value (Y ₁ Y ₂ …Y _k )
used in place of the standard pattern (X〓 ₁ X〓 ₂ …
Pattern matching with _X〓k ) is performed. According to this pattern matching, compared to the method of directly matching the measured value Y _i with the standard pattern, the fluctuation of the input audio is reduced in advance, so the separation between patterns is better, and the matching operation is faster and more efficient. It will be done.

FIGS. 3 and 4 are block diagrams showing a processing circuit for forming a time-series spectrum optimally estimated from the input speech.

In the figure, _M1 is _the first memory, which serves to store the standard pattern _, and the _vectors ( _{X〓 1} X〓 ₂ ... ing. M ₂ is the second memory and contains the measurement pattern (Y ₁ Y ₂ ...) introduced from the input section such as a microphone.
Y _k ), (U ₁ U ₂ ...U _k-1 ) are stored. The matrix A〓 ₁ A〓 ₂ ...A〓 _k-1 , which is a parameter of the dynamic model (time change of standard pattern), is stored in the memory _M4 , but in the embodiment shown in Fig. 3, the above parameter A〓 _i is In the case where the values are calculated sequentially, FIG. 4 shows the case where the predetermined values are already stored in the memory _M4 . Therefore, in the embodiment shown in FIG. 4, the memory _M4 stores the time series of matrices for each of the voice categories α, β, γ, etc. (conversion matrices for each category [A〓], [A〓], . , A〓 ₁ , A〓 ₂ ,…Ai〓…
,
A〓 ₁ , A〓 ₂ , ...A〓i ..., A〓 ₁ , A〓 ₂ , ...A〓i ...) are stored. In the embodiment of FIG. 3, the matrix A _i is sequentially calculated in the calculation unit C ₃ using the standard pattern stored in the memory M ₁ and is transferred and stored in the memory M ₄ each time. The main calculation unit _C1 is configured including a Kalman filter,
Equipped with the arithmetic circuit shown in Fig. 5, the above algorithm
Calculations based on equations (3), (4), (5) and (6) are performed.
Calculations are performed sequentially according to the time series, but when the calculation is performed at the i-th stage in the time series, each of P _i-1 and X^ _i-1 obtained in the previous stage The value is already stored in memory _M3 . At the initial stage, the memory M ₅ stores initial values P ₀ and X^ ₀ , respectively. Therefore, in the main calculation section _C1 ,
An estimated vector X^ _i of the feature vector is calculated at each stage using the values stored in the memories M ₁ , M ₂ , M ₃ and M ₄ , and the values are sequentially stored in the memories M ₄ and M ₅ .
will be stored in. That is, the estimated vector of the final frame is finally stored in the memory _M5 . In calculation section _C2 , the estimated vector X^ _i
The distance between the standard pattern feature vectors X〓 _i , X〓 _i . . . is calculated, added to the memory _M6 , and stored. D is a determination unit that compares the distances stored for each standard pattern in the memory _M6 and outputs a determination result. That is, it is checked to see which category the input voice belongs to.

As described above, according to the present invention, an optimally estimated feature vector is obtained from the measured speech using an evaluation function called a Kalman filter that takes into account the dynamic characteristics of the speech, and the feature vector is matched with a standard pattern. This reduces the fluctuation in the value of the feature vector compared to when obtaining the feature vector of the measured speech without using a Kalman filter, increases the efficiency of pattern matching, and improves the accuracy of separation between each speech category. can be improved.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing a dynamic model of a feature vector according to the present invention, FIG. 2 is a flowchart for explaining the operation of the present invention, and FIGS. 3 and 4 are for explaining an embodiment according to the present invention. Block diagram, No. 5
The figure is a block diagram showing the main parts of the same block diagram in detail. X _i : Feature vector, Y _i : Measured feature vector, M ₁ , M ₂ , M ₃ , M ₄ , M ₅ , M ₆ : Memory,
_C1 : Main calculation section, _C2 : Calculation section, D: Judgment section.

Claims (1)

[Claims] 1. For each category, the feature vector extracted from the audio signal is set in advance as a standard pattern of time changes that changes in a predetermined time series, and is extracted from input audio information. Evaluation function based on the time change of _the above standard pattern for measured _values X^ _i =X〓 _i +P _i W _i ^-1 (Y _i −X〓 _{i )} + _i-1 P _i = (M _i ^-1 +W _i ^-1 ) ^-1 M _i =A _i-1 P _i-1 A _i-1 +U _i-1 (X^ _i ; estimated feature vector, Y _i ; Vector by measured values, A _i ; Matrix describing time change, W _i ; Covariance matrix of measurement error, U _i ; Covariance matrix of excitation source,
P _i ：Covariance matrix of estimation error, _i ：Mean value of U _i ,
i; frame number) to estimate a feature vector of input speech, and compare the estimated feature vector with a standard pattern registered in advance to perform pattern matching of the speech signal. Method.