JPS61188599A - Voice recognition - 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 the Invention The present invention relates to a voice recognition method that allows a machine to recognize a human voice.

2. Description of the Related Art Speech recognition technologies have been actively developed and commercialized in recent years, but most of these are for specific speakers whose voices are recognized only by those who have registered their voices.

Devices for specific speakers require time and effort to register the words to be recognized in the device in advance, which puts a heavy burden on the user unless the device is used continuously for a long time. In response to this, research has recently been actively conducted on recognition technology for non-specific speakers that is easy to use and does not require voice registration.

Generally speaking, the speech recognition method performs pattern matching between the input speech and standard speech stored in a dictionary (these are parameterized), and selects the speech in the dictionary with the highest degree of similarity. The result of recognition is that the original direction is reversed. In this case, there is no problem if the input voice and the voice in the dictionary are physically exactly the same, but in general, even if the input voice is the same, different people say it or say it in different ways, so they may not be exactly the same. It won't be.

Physically, differences between people and differences in the way they speak are expressed as differences in spectral features and differences in temporal features.

In other words, the shape of the articulators (b, tongue, throat, etc.) differs from person to person, so the spectral shape of the same word will differ for different people.

Furthermore, the temporal characteristics differ depending on whether the vocalizations are uttered every day or slowly.

Speaker-independent recognition techniques require such spectra and their temporal variations to be normalized and compared to standard patterns. In the conventional example, nonlinear matching method (DP
This was done using the law). As an example of this, a matching method that applies the DP method to both spectrum and time is known (Miwa et al.: "Speaker-independent multi-word speech recognition using preliminary selection and spectral matching", Nippon Onkyo). Materials of the Society's Speech Study Group, 883-20.

19J33-6), so it will be explained as one of the conventional examples.

The input audio is analyzed every frame (10m5ec) using a filter of 29 channels, and this is zk (k: 1, 2
．． ...NUN:29). Further, it is assumed that the standard spectrum tri (t=112129) is used as a reference.

It is now assumed that the frequency fluctuation range is within 2 channels from tk+.

The distance between the input vector and the standard spectrum in nonlinear spectrum matching (1-C equation 1) k, ,
k are the beginning and end of the audio, respectively.

Here, c = (k+ l k, + 1+...
k2-1. 2) dks+ke=p(N (keJ,
'CkgJ)/(N-1ksl+N-1, l)
(Formula 2) The ramp function [ ] is defined as follows.

The denominator of (Equation 2) is the path normalization term in nonlinear spectral matching, and the numerator is the value at the end of the objective function calculated by the following recurrence formula.

(Formula 3) q(i, 1)=IZj-rtl (Formula 4) In word matching, the inter-word distance wi is calculated using the inter-spectral distance d obtained by (Formula 1).

The distance between words is also calculated using D'PF to deal with variations in speaking rate.
Calculate using j:.

w=h(I,J)/CI+J) ,, (Formula 5)
Here, I and J are the durations of the standard pattern and input pattern expressed in the number of frames, 1, and h(i, i) in (Equation 5)
is calculated using the following recurrence formula.

(2〇) Here, the frame number of the standard pattern and j is the frame number of the input pattern.

The word in the dictionary for which W in (Equation 5) is the minimum is output as a result. ”- In this way, in the conventional example, spectral fluctuations are normalized using (Equation 3), and time fluctuations are normalized using (20).

Problems to be Solved by the Invention As described above, in the conventional example, the DP method is used twice. Although the DP method requires a considerable amount of calculation even in the time axis direction (E6)), it is clear that if this method is used in the frequency direction ((Equation 3)), the amount of calculation becomes enormous.

The present invention solves the above problem, and aims to normalize both spectrum and time with a small amount of calculation, and to obtain a high recognition rate for speech of unspecified speakers.

Means for Solving the Problems The present invention expands and contracts the word length by dividing the input speech into a constant number of frames (J) between the start and end detected, and calculates all feature parameters for each of these frames. Calculation (number of parameters for one frame is 1d)
, n=ctx, constitutes a dimensional input pattern, and calculates the distance W between this and the speech standard pattern using a statistical distance measure, thereby absorbing spectral fluctuations and temporal fluctuations with a small amount of calculation. This enables the efficient recognition of voices from unspecified speakers.

Effects As mentioned above, the present invention replaces the spectral fluctuations and temporal fluctuations that are problematic in speech recognition of unspecified speakers with the characteristic parameter fluctuations and their temporal fluctuations, and transforms the form of the fluctuation into a multidimensional normal distribution. By assuming that the following is true, and incorporating it into the standard pattern as a statistic, it is possible to absorb the spectral shift between the unknown input and the standard pattern, and the frame shift when normalized over time. It is possible to realize a small, low-cost, non-specific speech recognition device that achieves a high overall recognition rate while reducing the amount of M.

The embodiment diagram is a functional block diagram showing an implementation of a speech recognition method in an embodiment of the present invention.

In the figure, 1 is a music converter that converts input audio into a digital signal, 2 is an acoustic analyzer that analyzes audio for each analysis section (frame) and obtains spectrum information, and 3 is a feature parameter extractor that obtains feature parameters. , 4 is a speech interval detection unit that detects the start frame and the end frame, 5 is a time axis normalization unit that expands and contracts the word length, 6 is a distance calculation unit that calculates the similarity between the input cover pattern and the standard pattern, and 7 is a distance calculation unit that calculates the similarity between the input cover pattern and the standard pattern. This is a standard pattern storage unit that stores standard patterns created in advance. In the above configuration, the operation vi'' will be explained below.

The input audio iAD conversion unit converts the input audio into a 12-bit digital signal. The sampling frequency is axuz. The acoustic analysis section 2 performs LPC analysis using the autocorrelation method for each frame (10 msec). The order of analysis is 1
+ order, and the linear prediction coefficient α. , α1. α2・・・・・・
α,. seek. In addition, the audio power W0 for each frame is also determined here. The feature parameter extraction unit 3 uses the linear prediction coefficients to extract LPC cepstral coefficients C1 to cd.
(d is censored next e,) and normalized logarithmic residual power Co
Find k.

Regarding the LPC analysis and the extraction method of LPC cepstral coefficients, see, for example, J.D., Markel, A.H.

This is described in detail in ``Linear Prediction of Speech'' by John Gray, translated by Hisaki Suzuki, so the explanation will be omitted here.

Further, the feature parameter extracting unit 3 obtains the logarithmic power LW0 using the following equation.

LW0=10 loq,. wo (Formula 7) The voice section detection unit 4 uses the LWO'e threshold θ6 determined by (Formula 7)
However, if the frames with LWo>θ6 continue for more than one frame, the first frame is set as the starting frame F of the voice section.
8. Also, after Fs, LWo and threshold θ. ″
f. When the frames in which Lyo<Oo are consecutive for 2 degrees or more, the first frame is the final frame F of the voice section. shall be. In this way, the period from Fll to Fe is defined as a voice section. Well, to simplify the explanation, let's consider it as the first frame of ys1 and use the frame number t"(' * 2+...i,...
・Engineering). However, I:=F, −F, +
It is 1.

The time axis normalization unit 5 linearly expands and contracts the word length by dividing it into J frame lengths. The i-th frame after expansion and contraction and the i-th frame of the input audio have the relationship (28).

, I-IJ-1 1=[end: irises-j+T ear+o, sJ (formula 8) but [
] represents the largest integer not exceeding that number. In the embodiment, J=16.

Next, the feature parameters after expansion and contraction are arranged in time series to create a time series pattern c8. Now, let the % signature parameter (Lpc cepstral coefficient) of the j-th frame be C(, (k: 0, 1
,2.・・・・・・e p: 4 pieces) and 1°, C becomes the following formula.

GX ” (Cj, OI C1, 11Cj, 2 ”””
Cj, p I ``'''' (Formula 9) That is, C becomes a J·(p++), that is, a J-d dimensional vector (d is the number of parameters per frame).

The distance calculation units calculate the degree of similarity between the input pattern and the standard pattern of each voice stored in the standard pattern storage unit 7 using a statistical distance measure, and select the voice with the smallest distance as the recognition result. Output. Assuming that the standard pattern 6cn (average value) corresponding to the n-th voice stored in the standard pattern storage unit is a covariance matrix 1ty common to all target voices, the n-th standard pattern is The Mahalanopis distance from the turn is calculated using the following formula.

Sn: (Gx-On)・W・(Cxi,) (Formula 1o
) The subscript t represents transpose, and -1 represents the inverse matrix.

Expanding (Formula 10) and removing terms unrelated to n, and assuming this as Dn, Dn can be calculated using the following equation.

D:b-alllc (Formula 11)%Formula% However, &Uni2・W-Cn CFormula 12)D4
is calculated for all n (n=1, 2...N), and the speech that minimizes Dn is taken as the recognition result.

Here, N is the number of voice standard patterns stored in the standard pattern storage section 7. Actually the standard number (turn is a
As many as the number of voices (N types) are stored, with n and bow as a pair.

The amount of calculation required to calculate (Equation 11) is that the multiplication is J・(p+1
) times, and the subtraction is once. In the example, J = 16, p
= 4, so by substituting this, the number of multiplications is 80
The conventional example ((Formula 3)).

(26)) It can be seen that the number is incomparably small.

Next, a method for creating standard patterns C and W (actually converted to &U, bn) will be explained.

A standard pattern is created using many data samples for each voice. For each voice, let M be the number of samples used. Apply (28) to each sample,
1, the number of frames is set to J, and the average value vector (Equation 14) is calculated for audio n. However, here, +rn in Cj1 is the m-th sample of audio n, and indicates the next Cepnutrum coefficient of the j-th frame. Find the covariance matrix of voice n using the same procedure as for the mean value vector. The covariance matrix ■ common to all voices is calculated using the following formula.

Placement=1(tw”+ w”+-+001.+ w”+ 0
00. −1゛1w ) (Formula 16) Cnli (Formula 12) and (Formula 13) a n *
bn and stored in the standard pattern storage section 7 in advance.

In order to confirm the effect of this example, 1o number (“ichi”)
, "two", "san", "yon", "go".

"Roku", "Nana", "Hachi", "Kyuu".

We conducted a recognition experiment on adult men and women regarding "Zero".

The table shows the relationship between the average recognition rate for all speakers and all words with respect to the truncation order p (number of parameters d=p+1) of the LPC cepnutrum coefficient.

According to the table, the highest value is reached when p = 4, and it can be seen that the recognition rate does not differ much beyond that. Therefore, p=4 is optimal from the viewpoint of reducing the number of dimensions used in the statistical distance measure and reducing the amount of calculation. The average recognition rate at this time is as high as 98.6%, and the amount of calculation is 800 times 1. Considering that the number of subtractions is very small, 10 times, the effect of this embodiment is obvious.

Although the explanation has been given using (Equation 11) as the statistical distance measure, it is also possible to replace this with other statistical distance measures such as the p\ize judgment and the a-type discriminant function. For example, if Bayesian judgment is used all around, the log likelihood Ln is obtained by the following equation.

Ln=-(Cx-Cn)@,-(Cx-Cn)-person.

(Formula 17) It is sufficient to output the audio in the standard pattern storage unit 7 that maximizes (Formula 17).

Furthermore, the feature parameters may be other than the LPC cepstrum, for example, the output of a bandpass filter, the PARCO cepstrum, etc.
R coefficients, autocorrelation coefficients, and modified versions thereof can be used.

Effects of the Invention In short, the present invention expands and contracts the word length by dividing the period between the start and end detected from the input audio into J frames of a constant length, and calculates feature parameters for each of these frames (per frame. (the number of parameters is d), construct a d×J-dimensional input pattern, calculate the distance between this and the speech standard pattern using a statistical distance measure,
The recognition result is the speech that corresponds to the standard speech pattern with the highest degree of similarity, and the spectral fluctuations and temporal fluctuations that have been problems in speech recognition for unspecified speakers are replaced by the fluctuations in feature parameters and their temporal fluctuations. It is assumed that the form of variation follows a multidimensional normal distribution, and it is incorporated into the standard pattern as a statistic. Therefore, the spectral deviation between the unknown human power and the standard pattern, and the frame deviation when normalized in time can be absorbed as long as the timing is statistically acceptable.

Furthermore, in the present invention, by not separating spectral fluctuations and their temporal fluctuations, but treating them at the same level (that is, treating them as temporal patterns of characteristic parameters), it is possible to reduce the amount of calculation, and reduce the amount of speech of unspecified speakers. It can be recognized with high accuracy even with a small amount of calculation, and its effects are large.

[Brief explanation of the drawing]

The figure is a functional block diagram embodying a speech recognition method in an embodiment of the present invention. 1... MuD conversion unit, 2... Acoustic analysis unit, 3... Feature parameter extraction unit, 4...
・Voice section detection unit, 6... Time axis normalization unit, 6
. . . Distance calculation section, 7. . . Standard pattern storage section.

Claims (3)

[Claims] (1) In advance, prepare as many speech standard patterns as the number of speeches to be recognized, which are composed of d×J orders of constant length J frames with d parameters per frame in the analysis section, and input them on the one hand. Detects the start and end of audio,
Divide the area between the start and end into J frames, extract d feature parameters for each frame, arrange them in temporal order to create a d x J-dimensional input time series vector, and combine this with each of the above-mentioned voices. A speech recognition method characterized in that the degree of similarity with a standard pattern is calculated using a statistical distance measure, and the result is that the speech corresponding to the speech standard pattern with the highest degree of similarity is determined as a recognition result. (2) The speech recognition method according to claim 1, wherein the statistical distance measure is Mahalanobis distance, Bayesian judgment, or a linear discriminant function. (3) The speech recognition method according to claim 1, wherein the feature parameter is any one of an LPC cepstral coefficient, an autocorrelation coefficient, and an output of a bandpass filter.