JPS617892A - Word speech 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 The present invention relates to a speech recognition method using vector quantization.

There is a method called the DP matching method (dynamic B [picture method]) for pattern matching between the standard pattern of the word "disaster" and the unknown cover pattern. this is.

Pattern matching is performed by non-linearly expanding or contracting the time axis of the standard pattern and aligning it with the time axis of the unknown input pattern so that both patterns are most similar, that is, the distance between the patterns is minimized. However, the above D
The P matching method requires at least I Requires.

Purpose The present invention was made in order to eliminate the drawbacks of the prior art as described above, and in particular, the frequency distribution pattern of the onomatopoeic pattern vector, which is obtained by vector quantizing the feature vector of the standard pattern, and the feature vector of the input pattern. In a speech recognition method that recognizes a word sound based on the shape between a vector quantized onomatopoeic pattern vector and a frequency distribution pattern, mutual similarity between pattern vectors is added to the pattern vector, and This was done with the aim of reducing the amount of calculation and improving the recognition speed.

The structure of the present invention will be described below based on one embodiment.

The figure is a block diagram of a system constituting the present invention. In the figure, 1 is an audio input section, 2 is a spectrum analysis section, 3 is a codebook storage section, 4 is an unknown input frame vector quantization section, and 5 is a code 6 is a standard frequency distribution pattern storage unit, 7 is a similarity table between code vectors, 8 is a frequency distribution pattern generator converted by the similarity table 7, 9 is a pattern matching unit , 10 is a word identification section, and ] 1 is a recognition result output section.

First, in the standard frequency distribution pattern storage section 6.

Let Rn represent the standard pattern of word n, and Rn = b,
n b, n , , , bn', , , , bn
Let J ・Jn (n=1.2..., NUN: number of words). Here, bn is the feature vector of the third frame of word n, and n is the continuous frame length.

Next, the standard pattern Rn is converted into an onomatopoeic silver pattern vector (also referred to as a code vector) included in the coat book 3.
It is expressed as Ck (k=1.2..., ni; ni: number of quantization levels). That is, the above b , n (, i=1..
2. ..., Jn) as the code vector C
It is expressed as the closest one among k.

Here, let the distance measure be d(b7.ch).

Facing, above, n(j= ], + 2 +...
・Let Jn) represent the standard pattern.

Let Y be the usage frequency of the code vector Ck for word n.
k', and the vector quantized standard pattern is
When expressed by the above Ykn as the 41 quasi-frequency distribution pattern R', it can be expressed as follows. Here, the code vector Cj (i = 1.2...,
■() In the similarity table 7 that reflects the similarity between the s
(++jL(i.

j=1.2. ..., K) is created in advance. As a measure of similarity, the distance d (Ci, Cj) between the code patterns 1 to C' and Cj is used, and d (Ci
, C, ]) is small, put a large value in s(+, j). For example, for a certain value of i, d (C''
, C, 1) + (j=], 2..., K) is the minimum (when ]=J, d(Ci, Cj)=O), then S(i.

, the elements of the similarity table S(i,j) are determined such that S(+1J)=5.

Next, the target $ frequency distribution pattern R' = Y.

¥-...'Yj Yk is calculated from the similarity table S(i, j
) to convert as follows. Mark 4 after conversion! It can be expressed as a frequency distribution pattern. The pattern of Rn is all words n (n=1.2
．． . . ., N) t: is determined in advance and stored in the mark frequency distribution pattern storage section 6.

On the other hand, the unknown input voice input to the voice input unit 1 is frequency-analyzed by the spectrum analysis unit 2 to obtain an unknown input/cover turn T. T can be expressed as T=a1a2...ai...aI. Here, ailt is the feature vector of the j-th frame, and ■ is the continuous frame length. .

The unknown input pattern T is also vector quantized by the code book 3 in the solid quantizer 4, and the pattern represented by the nearest code pattern (Ck) for each frame ai is assumed to be 1,000. , . =X,X,-...
...Xk...XI<. Similarity tape/L/7 (7) S l for this 1゛ as well as above.
, j) The pattern conversion section 8 obtains the converted result. If this is T, '1' is T=X,
X2...Xl...Xk.

Next, a pattern matching section 9 performs pattern matching with the frequency distribution pattern T of the standard frequency distribution information inputted to the pattern storage section 6. That is, the standard four patterns R
The word distance between n and unknown human pattern T is D (Rn
, T), and the target '$ frequency distribution pattern 12. ．．
Taking the frequency of use Yk of n and the unknown human power as df, the previous ia
The distance between single M'NID (Rn+T) is normalized by the sum of the input frame length ■ and the frame length Jn of word n, and is expressed as follows.

As the distance MA [df, the absolute value distance is used, and df ('Yn k, Xk) = l Yn k - Xk
Set to 1.

In addition, as the distance measure df in equation (1), the difference in value between the frequency number Xk of the knowledge input frequency distribution pattern T and the frequency number Xk is, for example, 1/
If it is in the range of 2 to 2 times, the distance scale df is set to 0.

It is also possible to use a distance measure. However, α.

β and γ are parameters, and α is the denominator of equation (2).
In order to avoid this, for example, α=1. β is a parameter for adjusting the distance scale, but normally β = 0, and γ is γ = 1
Use as.

By using the distance measure described above, pattern matching that is resistant to temporal nonlinear expansion and contraction of standard patterns and unknown input cover patterns can be performed. (2) Equation (
7) df(Y, "k, Xk) total xtt.

If you do this in advance and store it in a table, ~ to any Y', the distance df(Y
'k, Xk) can be immediately determined by quoting the table above.

Then, the inter-word distance D (Rn, T) is calculated for all dictionary words n (n=1.2..., N), and the word identification section 10 calculates the inter-word distance D (Rn, T). ) is output by the recognition result output unit 11 as the recognition result of the unknown input word.

That is, R is expressed as n=arg min D(R'+T).

Note that although an embodiment in which mutual similarity between code vectors is introduced using a similarity table has been described above, the present invention is not limited to the above embodiment. For example, for the first candidate code vector with a similar frequency, the frequency number (count number) is set to 2. For example, the frequency number (count number) may be set to 1 for the second candidate code vector to reflect the similarity between the code vectors.

As described above, according to the present invention, the number of calculations required for DP matching is KXN times (K<<
I It becomes possible.

As is clear from the explanations given below, according to the present invention, when pattern matching is performed based on the frequency distribution pattern of the feature vectors of word standard patterns and unknown input cover patterns, mutual similarities between code vectors are determined. By introducing this method, it is possible to improve recognition accuracy, reduce the amount of calculation required for pattern matching, and recognize unknown input speech words quickly and accurately. Furthermore, it has the advantage that it can be applied as a means for quickly and accurately preselecting big word meaning words.

[Brief explanation of the drawing]

The figure is a block diagram for explaining one embodiment of the present invention. 1...Audio input section, 2...Spectrum analysis section, 3...
・Codebook storage unit, 4: Unknown input frame vector quantization unit, 5: Code pattern usage frequency distribution pattern generation unit, 6: Standard frequency distribution pattern storage unit, 7: Code pattern 1- similarity table between rules, 8
. . . Frequency distribution pattern conversion section, 9 pattern matching section, 10. . . Word identification section, 11. . . Recognition result output section. Procedural technique [til three books (method) % formula %: 1. Display of the case 1982 Patent Application No. 12881.4 2. Name of the invention Word speech recognition method, person making the amendment Relationship with the case Patent applicant Ota ri Nakamagome Address 1-3-6 Nakamagome, Ota-ku, Tokyo Name (674) Ricoh Co., Ltd. - Representative Hama 1) Hiro (1 idiot), Agent Address 231 Naka-ku, Yokohama Furocho 1-2-
7 Shear Train Yokohama No. 807, subject of amendment (1), detailed description of the invention column 7 of the specification, content of amendment (1), page 3, line 7 of the specification, and page 11, line 12 1 in the description is corrected to ``Fig. 1 is''. (2) Add "Figure 1" to the diagram as written in red. 8. Attached documents report form 1st copy No. 11 [

Claims (1)

[Claims] Frequency distribution patterns of onomatopoeic pattern vectors obtained by vector quantizing feature vectors of standard patterns of words are stored in advance, and feature vectors obtained by spectrum analysis of unknown input word speech are also vector quantized to perform frequency analysis of onomatopoeic pattern vectors. A speech recognition method that performs pattern matching with a frequency distribution pattern of the word standard pattern represented by a pattern, characterized in that mutual similarity between pattern vectors is added to the pattern vector.