JPS60202500A - Phoneme recognition method - 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 phoneme recognition method in a speech recognition device characterized by performing phoneme recognition.

Configuration of conventional examples and their problems In recent years, a method has been proposed as a phoneme recognition method that uses the temporal movement of the spectrum as a feature parameter and applies a distance measure based on Bayesian judgment to this method, and good results have been obtained. There is a report that. Manate et al.: “Consonant recognition using time-frequency bangs” Acoustical Society of Japan Speech Study Group Material 19
(June 1982) The above-mentioned prior art will be explained below.

This report focuses on the fact that the characteristics of consonants exist not only in their frequency spectrum but also in their temporal movement, and uses the time-series information of the spectrum as a parameter. Spectral information is obtained by applying the audio signal to a 29-channel unimodal bandpass filter, and 29 spectral information is obtained every frame (1011 sec).
Pick up each item. This kl=cx4. x2...x2
．． ). Then, the time series vector 1J-(z,,.

2.129. j'1, 22, 229.2+, 31. ”'
"'J-29, 5, DC+, z"'"': 1:2g, y
l), and these are all feature parameters. For voiceless plosives and voiced plosives, m is set to 4. The following description will be made assuming m=4. If this is done, the number of dimensions will be 29×4=116, and the amount of calculation required for later similarity calculation will be enormous. Therefore, the number of dimensions of the parameters is compressed to 24 dimensions in two stages by principal component analysis.

First, in the first stage, it is compressed to 11: a total of 11 dimensions. The vector after compression is ==(11/, I2/..., 2
:11') and the coefficient matrix for compression is /
When A, the turtle f=lA-x(1). However, /A is a 29×11 matrix, which is prepared in advance from a large amount of data.

In the second stage, the data of 4 frames (11×4=44 dimensions) is further compressed into 24 dimensions by principal component analysis. The vector for 4 frames with respect to l' is J'-(,,
',,,x2',..."i'1,j"j',
2......"1'j, 21x+',!!""""1
'j, l"1', 4""""1'j, 4), and the compressed vector kfl=cy,,y2...y24)
Then, similar to (1), compression can be performed using the following equation.

y-Ya・K'(2) where l(d is a coefficient matrix having the number of dimensions of 44×24.

Next, phonemes are discriminated using the parameters compressed in this way. As a standard pattern of phonemes, the standard pattern prepared in advance is (1), (2) for a lot of data.
) and calculate their mean and variance-covariance. Here j is the phoneme number, for example 1:/c
/, 2: /p/.

3: /l/, 4: /ni/. The degree of similarity Pj of unknown human power V to the phoneme with phoneme number j is determined by the following equation.

Here, the subscript T means transpose, and -1 means inverse matrix. (3
), take the logarithm and double it, and let this be Lj6, then Lj-(W-uij)T・Σj・(y-pj)+C, (
4) Here, Cj-21og((2π)12・lΣjI)
becomes.

The phoneme corresponding to the phoneme number with the maximum value (4) is set as the unknown person's discrimination result. (Bayesian Determination) The problem with the above conventional example is that the amount of calculation is large. In the conventional example,
) and (2) compress the 116 dimensions to 24 dimensions, and (
3) The amount of calculation required for similarity calculation is reduced. However, instead, calculation of the amount of compression is required.

OBJECTS OF THE INVENTION The present invention eliminates the above-mentioned drawbacks of the prior art, and provides a phoneme recognition method that requires less calculation and is capable of obtaining good recognition results.

Structure of the Invention The present invention achieves the above object, and the basic structure of the present invention is to use time series information of LPC cepstral coefficients as a feature parameter, and to apply a Bayesian 6-bench algorithm to this.

This system uses statistical distance measures such as the distance and Mahalanobis distances to discriminate between phonemes, and by truncating the order of the LPG cepstral coefficients below the analysis order, the amount of calculation can be significantly reduced without reducing the recognition rate. It has been reduced.

Description of examples Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 shows a block diagram for implementing a phoneme recognition method in an embodiment of the present invention. In the figure, reference numeral 1 denotes an acoustic analysis unit that converts input audio into a digital signal, performs LPG analysis for each complaint at the N-order analysis order, and converts the input audio into a digital signal.
Calculate the LPG cepstrum coefficient of ≧n (including coi)
. Reference numeral 2 denotes a parameter storage unit that stores parameters (LPG cepstrum coefficients) for m frames. Reference numeral 3 denotes a phoneme standard pattern storage unit in which characteristic parameters are determined in advance from the voices of many speakers, and standard patterns created for each phoneme based on these characteristic parameters are stored. In the 4id similarity calculation unit, mX(n+1
) unknown parameters and the standard patterns of each phoneme stored in the phoneme standard pattern storage section 3. 5 is a phoneme discriminating unit which outputs all phoneme numbers with the maximum similarity as recognition results.

Below are examples of voiceless plosives (/p/, /l/).

/, /, 10/) We will describe a method for phoneme discrimination using k objects.

The input speech is analyzed frame by frame by linear predictive analysis (LPG analysis) in the acoustic analysis unit 1, and LPG cepstral coefficients are calculated. The window used for analysis is a 20 m5ec Hamming window, the sampling frequency is 12 KHz, and the analysis order is 15F (autocorrelation method). Regarding LPG analysis and calculation of LPC cepstral coefficients, for example, J.
Dl Markel.

Written by A. H. Gray, Jr., translated by Hisaki Suzuki, "Linear Prediction of Speech", 1986, published by Corona Publishing.

LPG cepstrum coefficient '6 obtained for each frame
cm (co, cl...On), however,
co is the power ring of LPG cepstral coefficients (corresponds to the 0th order term).

Time series vector 1llll-(CO
,1"1.1"2.1"...'n,1,CD,2,C1
, 2"'"'On, 21CD, 51CI, 3"""On
,3”””CO,m”1. m'...""n, m). ID1d(n+1)/m-dimensional vector. This is stored in the parameter storage section 2. The similarity P for the phoneme of phoneme number j in the time series vector is determined by the following formula in the similarity calculation unit 4: % Formula % Here, the subscript T means transposition, and -1 means an inverse matrix.

If we take the logarithm and double it in (5) and let it be L-J, then L・-1(μj)・Σj(tl11j)+C, (6)
Here, Cj-21og((2π)12·lΣj11.

Therefore, by creating a pattern using μj, Σ, and 1LPc cepstral coefficients, storing it in the phoneme standard pattern storage unit 3, and calculating the maximum value of 7, the phoneme discrimination unit 5 can discriminate the phoneme.

The feature of this embodiment is that the number of dimensions of the time series vector can be reduced by simply truncating the order of the LPG cepnutrum coefficients without using a method such as principal component analysis.

Next, constants that define the number of dimensions of (1) - the number of frames m and the truncation order n- will be described.

First, Table 1 shows the relationship between the number of frames m and the phoneme discrimination rate.

However, in this case, the truncation order was set to n=4.

(Margin below) 10 be.

As can be seen from Table 1, the discrimination rate for each phoneme is saturated at about 3 frames, and the average discrimination rate for each phoneme is also saturated at 3 frames. Although 75f0.2% is higher for 4 frames, this difference is not a problem from the viewpoint of compressing the number of dimensions. Therefore, the number of frames m=3 is appropriate.

Next, the number of frames m is fixed at 3, and the truncation order n of the LPG cepstrum coefficients will be described. FIG. 2 illustrates the relationship between the truncation order and the phoneme discrimination rate, where the horizontal axis is the truncation order n, and the vertical axis is the discrimination rate (unit: %). The thin solid line is /C/, the broken line is /p/, the one-dot chain line is 7t/, and the two-dot chain line is /ni/.

The average discrimination rate of all phonemes is shown by a thick solid line. As is clear from the figure, the average discrimination rate is almost saturated at n = 4, reaches the highest value of 88.1% at n = 5, and after n = 6 /
As p/ decreases, the average discrimination rate also appears to gradually decrease. In this way, it has been found that the order of the LPG cepstral coefficients may be approximately 6th order, and higher orders are not required. The reason lies in the nature of cepstral coefficients. Since the cepstral coefficients are the inverse Fourier coefficients of the logarithmic spectrum,
The low-order coefficients represent global features of the spectrum, and the high-order coefficients represent detailed features of the spectrum. The detailed features of the spectrum are easily influenced by variable factors such as individuality and the environment in which they appear, and from a statistical perspective they do not significantly contribute to phoneme discrimination. Therefore, what is effective for phoneme discrimination is the global feature of the spectrum, and it is sufficient to truncate the cepstral coefficients at a necessary and sufficient order to express this feature.

In the case of voiceless plosives, the order is 4th to 6th, and the same is true for voiced plosives and nasals. On the other hand, in semivowels, the truncation order is 7th to 8th. In general, it is sufficient that the truncation order n of the LPG cepstrum coefficients is less than or equal to the analysis order of LPG analysis, and in most cases,
/3 or less is sufficient.

In this way, in this embodiment, LPG is used for parameter compression.
It is only necessary to truncate the order of the cepstral coefficients, and no calculation is required.

Next, voiceless plosives (/p/, /l/, /ni/.

10/)'e As an example, the amount of calculation will be compared with the conventional example.

It is assumed that the truncation order n of the 13th LPC cepstral coefficient is the 6th order (ie, C6, C4, . . . C5), and the number of frames is mi3. Therefore, the number of parameters is 6×3=18.

Table 2 shows a comparison of the amount of calculation with the conventional example.

However, Cj was treated as a constant using similarity calculation I/1 (8).

(Margin below) 14..

15=,= As can be seen from Table 2, this example does not require calculation for parameter compression, and since the number of parameters is smaller than the conventional example, the amount of calculation required for similarity calculation is also the same as that of the conventional example. It should be much less than that.

The total amount of calculation required to discriminate voiceless plosives can be reduced from 11/2.7 multiplication of addition/subtraction to 1/2.8 in the conventional example. As described above, in this embodiment, the amount of calculation can be considerably reduced compared to the conventional example, and the phoneme discrimination rate remains almost the same. Comparing the average phoneme discrimination rates, the conventional example is slightly better for voiceless plosives, 88.1% in the present example and 90% in the conventional example, but on the contrary, for voiced plosives, the present example:
86.4%.

It is 86% of the conventional example, and this example is superior. Overall, it can be seen that there is almost no difference in the discrimination rate. The fact that it does not reduce the phoneme discrimination rate and requires less calculations means that
There are great advantages in that the device can be made smaller and cheaper.

Note that in the above explanation, the distance based on the Bayesian judgment shown in (6) is used for the similarity calculation in the similarity calculation unit 4, but other statistical distance measures may be used. For example, the mano-lanobis distance Mj is expressed by the following equation.

M, −(El-ptj)・Σ・〇−111j)=“・Σ
-1・])-2piT-X'・F+,T,x1・μノ(
7) Since the first term in (7) is unrelated to j, it is not necessary when comparing only the sizes. The third term is a constant unrelated to the unknown input, and will be expressed as Fj. Therefore, (7) can be simplified as follows.

Mj'=Fj-&j-V However, aj=2・ptj・Σ(8) (8) is characterized in that the amount of calculation is even smaller than (6). Therefore, even when (8) is used, this embodiment is even more advantageous than the conventional example.

Of course, a linear discriminant function can also be used as another statistical distance measure.

Effects of the Invention To summarize, the present invention uses time series information of LPG cepstral coefficients as a characteristic parameter, and uses the above-mentioned 17.

This method calculates the degree of similarity with a pattern using a statistical distance measure to discriminate all phonemes, and the order of the LPG cepstral coefficients'! This provides a phoneme recognition method that aborts the analysis order below the analysis order of 1l-LPG analysis.Compared to conventional methods, the amount of calculation can be significantly reduced while maintaining performance, and the device can be made more compact. It has great practical advantages such as low cost.

[Brief explanation of the drawing]

FIG. 1 is a block diagram that embodies a phoneme recognition method according to an embodiment of the present invention, and FIG. 2 is a diagram showing the relationship between the discrimination rate of voiceless plosives and the truncation order of LPG cepstral coefficients. 1... Acoustic analysis section, 2... Parameter storage section, 3... Phoneme standard pattern storage section, 4.
. . . Similarity calculation section, 5 . . . Phoneme discrimination section. Name of agent: Patent attorney Toshio Nakao and one other person

Claims (2)

[Claims] (1) Using the time series information of t and pc cepstral coefficients as feature parameters, prepare in advance a phoneme standard pattern for each phoneme or phoneme group from the voices of many speakers based on this feature parameter, and When linear prediction (Lpc) analysis is performed with the following analysis order to determine the time series information of the n-th I and PC cepstrum coefficients as a feature parameter, the order of the LPC cepstrum coefficient nf is equal to or less than the analysis order N of the LPO analysis, and the unknown A phoneme recognition method characterized by discriminating phonemes by calculating the degree of similarity between input speech and a standard phoneme pattern using a statistical distance measure. (2) The phoneme recognition method according to claim 1, characterized in that the statistical distance measure is any one of Bayesian distance, Ma-Lanobis distance, and linear discriminant function. 2.、ノ