JPS62116999A - Syllable unit voice 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] [Field of Application of the Invention] The present invention relates to a speech recognition device, and particularly to speech recognition devices for large words, to which it has been difficult to apply word-by-word speech recognition devices, which have been commonly used in the past. The present invention relates to a monosyllable unit speech recognition device suitable for speech recognition.

[Background of the invention]

In the conventional monosyllable unit recognition method, large word speech recognition using monosyllable unit input from unspecified speakers (Transactions of the Institute of IEICE, V
ol, J 65-D, &12) As in (7), vowels that are relatively stable in a single syllable are first detected, and consonants and vowels are twisted in order to deal with deformation due to articulatory combination.

Alternatively, it was common to recognize chains of consonants, vowels, and consonants using methods such as DP matching.

However, this method requires segmentation of consonant and vowel intervals, which is difficult, and because matching is performed based on a single distance information, essential but minute features are lost. The problem was that the recognition rate did not improve.

[Purpose of the invention]

The purpose of the present invention is to avoid the difficult problem of segmentation of vowel intervals and consonant intervals, and to create a single syllable with a high recognition rate that can be used finely without losing the essential features, although they are weak. The object of the present invention is to provide a unit speech recognition device.

[Summary of the invention]

Input speech is converted into a series of spectra (approximately 10-dimensional vector representation) by acoustic analysis. The spectral series is
They are quantized by comparison with standard patterns and converted into symbol strings called primitives. The methods described so far have been used in conventional recognition methods.In the present invention, the conventional methods of segmentation and matching are not used in the recognition process of primitive sequences, but recognition is performed based on generative grammar. do. Each monosyllable (l
al=1Kal, etc.) is expressed as a sequence of primitives. Therefore, rules for generating primitive sequences for each monosyllable are determined in advance. Monosyllables can be recognized by analyzing which production rule is used to generate the input primitive sequence.

[Embodiments of the invention]

Embodiments of the present invention will be described below with reference to FIGS. 1 to 3.

The entire procedure for monosyllable recognition is shown in Figure 1. The input audio 6 is
It is processed by the acoustic analysis unit 1 and converted into a series of spectra. The spectral sequence is compared with the primitive standard pattern 3 by the quantization unit 2 and converted into a primitive sequence. The sequence of primitives is analyzed by the parser 4 using production rule 5 and output as a monosyllabic sequence 7.

The analysis by the parser 4 will be explained with reference to FIG.

The sequence of primitives input to the parser is shown in Figure 2 (a
) is a symbol string as shown in Figure 2 (a) is 1Kal
An example of a Cv syllable is shown, which consists of a silent part that occurs between the preceding syllable, a consonant part, and a vowel part. Among these, silent parts and vowel parts are expressed as arbitrary lengths because they are expanded or contracted in time depending on the speech situation. The production rule for this monosyllable is shown in FIG. 2(b). Here, S is the start symbol.

Capital letters such as A and B are non-terminal symbols (rewritable variables), and lowercase letters such as y a+ b are terminal symbols (primitives). By applying the production rule shown in FIG. 2(b) to the symbol string shown in FIG. 2(a) and replacing terminal symbols with non-terminal symbols, the starting symbol S is finally obtained. At this time, the symbol string follows the grammar prescribed by the production rule, and is recognized as a monosyllable corresponding to the production rule.

Since the symbol string is actually input continuously,
The time when parsing (analysis using production rules) for one single syllable ends is the time when parsing for the next single syllable begins. Therefore, it is not necessary to segment monosyllabic sections, vowel and consonant sections, etc. in advance.

Production rules are created for each syllable, that is, for each syllable. Variations in changes in symbol strings due to articulatory combinations are also embedded in the production rules. In addition, regarding the parsing method, in order to prevent symbols from being dropped or inserted due to noise or speech conditions,
It is desirable to use an error-correcting purging method.

FIG. 3 shows an embodiment of the word speech recognition device according to the present invention. 4, 5, and 7.9 are devices for acoustic analysis, quantization, purging, and matching, respectively. Also, 6
, 8.10 are a memory storing standard patterns of primitives, a memory storing production rules, and a memory for a word dictionary, respectively. The input speech 11 is first converted into a spectral sequence in step 4, and the spectral sequence is converted into a primitive sequence in step 5. At 7, the primitive sequence is parsed and recognized as a monosyllable.

The monosyllable sequence is matched with the word dictionary in step 9 and is recognized as a word. 3 is a bus for communication between devices, and the controller 1 performs synchronization control and bus control between the devices. The recognition result is used by the host computer 2.

〔Effect of the invention〕

According to the present invention, recognition is performed in units of monosyllables, so it goes without saying that even when the number of words increases, the effort required for template creation and matching does not increase. When compared with conventional methods as a recognition method for This has the effect of making it possible to make detailed use of weak but essential features that would otherwise be lost when performing a process, and is expected to improve the recognition rate.

[Brief explanation of drawings]

Figure 1 is a procedure diagram for monosyllable unit speech recognition, Figure 2 (a)
is an example of a primitive sequence, FIG. 2(b) is a generation rule corresponding to FIG. 2(a), and FIG. 3 is an example of a word speech recognition device. Tomi 1 Figure Tomi Z Figure (L) rut -1'l l"+Qz4ab+ -b, bt-
bz(b)

Claims (1)

[Claims] 1. A means for acoustically analyzing input speech, a means for quantizing a vector sequence obtained as a result of the acoustic analysis, and a method for recognizing monosyllables by analyzing the symbol string obtained as a result of the quantization using production rules. A syllable-by-syllable speech recognition device characterized by having means for performing.