JPS6026399A - Word recognition equipment - 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 improvements in word speech recognition devices.

Conventionally, there has been a method of matching words in which the words to be recognized in speech recognition are registered as standard patterns, the unknown words input during recognition are compared with the standard patterns, and the one with the highest degree of similarity is determined as the recognition result. Widely used.

This method can be applied not only to specific speakers but also to unspecified speakers. In other words, by preparing a plurality of standard patterns for each word, it is possible to cope with variations in the slam patterns depending on the speaker. These standard patterns can be determined as representative patterns from a large number of speakers using a clustering method.

When recognizing multiple words input by a certain speaker using such multiple standard patterns, the first input word is most similar to the standard pattern R1, and the next input word is the standard pattern R1. The most similar to the horse is the standard bang bird and R
A case may arise where 3 is the utterance of a completely different type of speaker. Although such unreasonableness leads to erroneous recognition, such unreasonableness can be avoided if it is known that the plural words to be recognized were uttered by the same speaker.

An object of the present invention is to provide a speaker-independent word recognition device that can avoid erroneous recognition due to differences in speakers and improve recognition accuracy.

The present invention includes a word detecting means for detecting a word section in an input audio signal, an input drum storage section for temporarily storing the drums of the detected word section for at least two consecutive words; a standard button storage unit that stores standard buttons prepared for each word;
The recognition unit includes an identification unit that performs matching between the manual slam read out from the manual slam storage unit and the standard pattern read out from the standard slap storage unit each time a word is detected, and obtains a recognition result.

The principle of recognition in the present invention will be described below. In the present invention, recognition is performed in units of word concatenation, in which inputted words are grouped into a plurality of consecutive words. The word patterns of multiple speakers prepared to create standard patterns are A.
p (represented by Wl. Here, p represents the speaker and W represents the word. From these word patterns, the two-word concatenation button Ap (W, , W2) = Ap (Wl) ■Ap
Add (%). Here, the symbol ■ indicates that two word patterns are juxtaposed.

For each combination of word connections (w,, w2), the slam set of multiple speakers (A-p(Wl, Wt) * p= L2...
・ From ), a standard batan set of word connections (Rk(w, , w2), k= 1.2...
). As a clustering method, k-mean
Any method such as the s method can be used.

Next, the principle of recognition in the present invention will be explained with reference to FIG. In FIG. 1, (a) shows a speech waveform, (bj is a word pattern, and fc) is a word concatenation pattern consisting of two connected words.

The input multiple word patterns are ``!+...S3...
・.

8. Let it be T. From now on, word concatenation slam S! ,...,
Sj.

...*5J-1 is generated as follows.

Book S = 8. ■52 8j 工S'' ■Sj+ is matched against each word concatenation button using a word concatenation standard button set. That is, D(Sj
, Rk(vl, w2)) for 0w2, D
W2 when is the minimum is the recognition result. However, here, D(A, B) represents the distance between the two buttons A and B, and v4 represents the recognition result of the third word.

Since the word concatenation for standard baton creation is a concatenation of words from the same speaker, matching based on word concatenation as a unit also assumes that the words in the word concatenation are from the same speaker. Since word concatenation is generated by shifting input words one by one, it is assumed that all input words are from the same speaker.

According to the present invention, since the identity of speakers for a series of inputs is assumed at the time of recognition, it is possible to realize a highly accurate speaker-independent word recognition device.

An embodiment of the present invention will be described below with reference to Figure 2.

FIG. 2 is a block diagram showing one embodiment of the apparatus according to the present invention. The input audio signal is processed by the word detection unit 1.
A word section is detected, the signal within the section is analyzed, and the input word bang is output. As the analysis method, any method such as a group of band-pass filters, linear predictive analysis, or fast Fourier transform can be used, and any method such as a method using energy can be used as the word section detection method. The format of the button is a time series of feature vectors obtained by analysis, but it may also be a format in which this time series is dimensionally compressed or time axis compressed. The input button storage section 2 is composed of two shift buffers 21 and 22. - When a word is detected, the contents of the buffer 21 are transferred to 22, and 21 stores the button of the newly detected word. Storage part 3
is used to store word concatenation standard buttons. Identification part 4
Each time a word is detected, the input word concatenation button read out from 21 and 22 is matched with the standard button read out from the storage section 3, and a recognition result is output. Reference numeral 41 is a matching unit that performs matching between the input baton and the standard baton and calculates the degree of similarity. Here, any matching method may be used as long as the degree of similarity can be calculated.

Reference numeral 42 denotes a matching control section, which specifies the previous recognition result l held in the recognition result output buffer 43 and a standard button to be read out from the storage section 3, and also controls the matching section. For example, when the j-th input word is detected and the recognition result of the content of 43 is Vj, the standard pattern of word concatenation (Wl.

Only those in which W2) is (vj, w2) are matched, equation (1) is calculated, and the result is output to 43.

As explained above, according to the present invention, it is possible to realize a word recognition device for non-specific speakers with high recognition accuracy.

Further, in the above description, a case has been described in which recognition is performed in units of two-word concatenation, but it is also possible to use three or more word concatenations. In that case, the number of shift buffers included in the storage section 2 changes. For example 3
In the case of word concatenation, the storage unit 2 is composed of three shift buffers.

[Brief explanation of the drawing]

FIG. 1 is a diagram for explaining the principle of recognition, and FIG. 2 is a block diagram showing an embodiment of the apparatus according to the present invention. In the figure, ■... word detection section, 2.3-- storage section, 21, 22... shift buffer, 4... identification section, 41... matching section, 42... matching control section , 43... Output rose "tl" Figure 2

Claims (1)

[Claims] word detecting means for detecting a word section in an input audio signal; an input slam storage section for temporarily storing a slam in the detected word section for at least two consecutive words; A recognition result is obtained by matching the input button read from the input button memory section and the standard pattern read from the standard button memory section each time a word is detected. What is claimed is: 1. A word recognition device comprising: an identification unit that obtains a word recognition unit;