JPS59168498A - Continuous voice recognition system - 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 continuous speech recognition method, and more particularly to a continuous speech recognition method that enables recognition of continuous speech simply by registering a small number of standard patterns.

[Prior art]

Conventionally, in a speech recognition device that recognizes speech that is naturally and continuously uttered at the character level, each individually discrete utterance has been registered and used as a standard pattern.

However, since the speech pattern during continuous utterances differs from the speech pattern during discrete utterances due to phenomena such as articulatory coupling and VC, the standard pattern using discrete utterances cannot be said to be qualitatively satisfactory.

In addition, in order to recognize speech with arbitrary content, it is necessary to recognize phonemes and syllables, but for this purpose, the recognition unit is short and the unit that takes into account the effects of deformation due to articulatory combination, etc. It is desirable to register standard patterns such as VC ■ phoneme chains such as vowel-consonant-vowel.

However, the above VcV phonological chain unit (hereinafter simply r
"vcv unit") is at least 7 in Japanese.
There are more than 00 types, and there is a problem in that the burden on users to register is extremely heavy.

[Purpose of the invention]

The present invention has been made in view of the above circumstances, and its purpose is to solve the above-mentioned problems in the conventional continuous speech recognition method, and to recognize continuous speech by simply registering a small number of standard patterns. The purpose of this invention is to provide a continuous speech recognition method that enables continuous speech recognition.

[Summary of the invention]

The key point of the present invention is that in a continuous speech recognition method that performs speech recognition by comparing with pre-registered standard patterns, only a limited number of standard patterns that appear frequently in practical use are registered, and Regarding patterns, approximate patterns are created and registered by combining partial patterns of the standard patterns that appear frequently and are registered above.

In addition, if the recognition result using the above approximate pattern is poor, a good pattern can be created by replacing the partial pattern in the corresponding input cover turn based on the information during error correction. It is also possible to change it.

The principle of the F1 invention will now be explained in detail.

Figure 1 shows the cumulative frequency of the VCV units that appear in the data, arranged in order of frequency of use, based on data surveyed by the National Institute for Japanese Language and Linguistics on the frequency of use of the word millet in 90 contemporary magazines. It is. Here, V (vowel) is 1a1,
1 I 1, 1 1, 1 G 1゜ 1 first 5 vowels and pellicle IN1
, consonant IQ+, and non-speech intervals outside the beginning and end of words are treated as vowels.The total number of types in p, vcv units reaches approximately 780 types.

By the way, according to Figure 1, about 780 types of VC
It can be seen that the top 200 most frequently used types of V cover approximately 89% of the total. Therefore, these top 200 types■
The CV unit is registered as a standard pattern, and the other CV units are created and registered to use approximate patterns by combining partial patterns of the registered standard pattern in the apparatus. The above combination is, for example, to synthesize the pattern of l eku 1, which is a VCV pattern of infrequently used, from l kul of l akul, which is frequently used, and 1 ekI of 1eki 1, which is also frequently used.

[Embodiments of the invention]

Embodiments of the present invention will be described in detail below with reference to the drawings.

FIG. 2 is a block diagram of a speech recognition device showing one embodiment of the present invention. In the figure, 2 is a voice analysis section, 3 is a backup memory, 4 is a standard pattern memory, 5 is a pattern matching section (hereinafter simply referred to as "matching section"), 6
7 indicates an input display section, 7 a main memory, 8 a control section, and 9 a flexible disk memory. The above-mentioned components themselves are conventionally known, and the main memory 7 stores the above-mentioned frequently appearing standard patterns (hereinafter referred to as "highly used VCV information").

The operation of the apparatus of this embodiment will be explained below.

First, voice registration of frequently used VCV information to be input is performed as follows. Under the control of the control section 8, the user is instructed through the guidance section of the input display section 6 to utter the VCV to be input according to the frequently used VCV information in the main memory 7.

When the user inputs the designated VC ■ voice 1 in accordance with the above instructions, the analysis section 2 performs feature extraction, and inputs it into the main memory 7 via the input cover sofa memory 3 . The input pattern is cut out according to the instructions of the control unit 8 in a necessary range as a standard pattern, and is registered in the standard pattern memory 4.

When all of the frequently used patterns have been input and registered according to the above procedure, the control section 8 creates a lowly used pattern by combining one division of the already registered frequently used patterns and registers it in the standard pattern memory 4.

The above-mentioned division is based on the fact that the power of a vowel is generally lower than that of a vowel, so by detecting the temporal dip position of the power, the VCV pattern can be easily divided into VC and CV. Each standard pattern created as described above can be stored in the flexible disk memory 9 and reused at a later date, if necessary.

Recognition is performed as follows. Input 1 is subjected to feature extraction in analysis unit 2 and stored in temporary buffer memory 3, and then
In the matching section 5, the pattern is matched with each of the standard patterns. In order to match continuous input audio with the standard pattern while detecting the corresponding interval with the standard pattern, for example, the present applicant first proposed Utility Model Application No. 1983-66296 ``
A device such as that proposed as a "speech recognition device" can be effectively used.

The above matching result is taken into the control section 8.

The control unit 8 recognizes a phoneme string that constitutes the input speech from the matching result, and displays it on the input display unit 6 as a character string. The user looks at the displayed results, and if there is an error, inputs correction information through the input display section 6.

When there is an error correction input, the control section 8 cuts out the corresponding portion in the input buffer memory 3 and replaces it with the registered pattern.

Here, the recognition rate (proportion of correct recognition) when using the officially registered standard pattern is over 95%, while the recognition rate when using the approximate pattern by combining two divisions is 85%. It has been experimentally confirmed that the The ratio of vCv patterns that have been erroneously recognized and should be newly registered as standard patterns can be approximately 90 types in terms of degree and number.

Furthermore, if an error occurs, the standard pattern is corrected using the error correction information while being used. If combined with the method proposed by the present applicant in Japanese Patent Application No. 57-197511 "Speech Recognition Device", it is possible to register a good standard pattern relatively easily by sequentially approximating it while using the device. is possible.

In addition, even with only the above 200 standard patterns at the time of initial registration, the recognition rate considering the frequency of appearance is 0.89 x 0.9
Needless to say, it is possible to satisfy 5+0.11X0.85=0.94 (%).

In addition, if there are no 8-figure partial patterns in the initial registered standard pattern to create other patterns, you may need to register only the monosyllable C■ separately, add some VCv, or For some VCs, it is also possible to substitute other patterns having C of the same articulatory style. It is also possible to replace the pattern generated by the above-mentioned substitute pattern with the input cover pattern.

In the above example, the vowel sounds and consonants were included in the vowels, but
It goes without saying that it is possible to handle these separately and to perform modifications such as handling VNCV and VQCV patterns.

〔Effect of the invention〕

As described above, according to the present invention, in a continuous speech recognition method that performs speech recognition by comparing with standard patterns registered in advance, only a limited number of standard patterns that appear frequently in practical use are registered. , for patterns that appear less frequently, an approximate pattern is created and registered by combining the subpatterns of the patterns that appear more frequently registered above, reducing the burden on the user when registering standard patterns. This has the remarkable effect of being able to realize a continuous speech recognition method with reduced noise.

[Brief explanation of the drawing]

FIG. 1 is a graph showing the cumulative appearance frequency of units of ■C■ in continuous speech, and FIG. 2 is a block diagram of a speech recognition apparatus showing an embodiment of the present invention. 1... Input, 2... Voice analysis section, 3... Backer memory, 4... Standard pattern memory, 5... Matching section, 6... Input display section, 7... Main memory ,
8...Control unit, M 1 Port ζ 2 Deposit 2 Figure

Claims (1)

[Claims] At least a voice analysis section and a standard pattern storage section. In the continuous speech recognition method, which has a pattern matching section and an input display section, and performs speech recognition by matching with standard nodules registered in advance in the standard pattern storage section, the method is stored in the standard pattern storage section. Standard patterns are limited to those that appear frequently in practice.
For other patterns that appear less frequently, create/approximate patterns that combine the above registered patterns or parts of them.
A continuous speech recognition method characterized by registration.