JPS58220197A - Standard pattern preparation system for recognition of simulated continuous utterance - 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] (a) Technical Field of the Invention The present invention relates to a speech recognition device that recognizes input speech using pseudo-continuous speech (pronounced slowly to bring out the characteristics of syllables), and particularly relates to a In order to reduce the burden on the operator, the standard patterns to be registered in advance are made using only monosyllables.
The creation and learning of standard patterns for improving the recognition rate of input by pseudo-continuous utterances relates to a standard pattern creation method in automatic pseudo-continuous utterance recognition.

(b) Prior art and problems In the conventional technology, it is practically impossible to input sentences, etc. into a speech recognition device in monosyllable form because the user's utterances are unnatural. The voice input must be able to occur continuously as much as possible. In this case, it is natural that the acoustic characteristics of each phoneme are affected by the preceding and preceding phonemes due to continuous utterances, and the recognition ability is affected by problems such as articulatory coupling and problems such as the characteristics of each phoneme not being clearly visible. However, it is not necessarily possible to separate monosyllable units, so even if you register only monosyllables as standard patterns in advance, pattern matching with monosyllables cut out from pseudo-continuous utterances will not be possible. You won't get good results. Therefore, it is desirable to register not only a standard pattern consisting only of middle syllable (consonant + vowel) patterns, but also vowel and monosyllable patterns as standard patterns and perform pattern matching. However, the Kishi syllable is (consonant → −
Even if a vowel (vowel) is uttered - times and registered, there are more than 100 types, and uttering and registering vowels and monosyllabic patterns burdens the user more than six times.
Moreover, it has the disadvantage of being incomplete.

(c) In order to lighten the purpose of the invention and improve the recognition rate of input using pseudo-continuous utterances, the standard pattern to be registered in advance is only a single syllable, and the utterance of the single syllable is limited to each time. , vowel and monosyllabic patterns are used as standard patterns to increase the recognition rate of pseudo-continuous utterances. However, the vowel and monosyllabic patterns are organized into a specified number of patterns from the vowel and monosyllabic syntax 1 information created in advance, and the axillary continuous utterances are redivided based on the results of the post-processing device. An object of the present invention is to provide a standard pattern creation method that obtains learning pattern sections, performs learning, and reorganizes the patterns to create vowel and monosyllable standard patterns.

In a speech recognition device, only a single syllable is used as a standard pattern to be registered in advance, and several types of standard patterns are generated for each monosyllable, and patterns are also created in advance for each type of syllable. Based on the statistical information of vowels and monosyllables, the speech input by pseudo-continuous utterance is divided into I parts to obtain learning pattern sections, and the pattern is reorganized into vowels and monosyllabic standard patterns. It was designed to do so.

(e) Embodiment of the Invention Based on the statistical information created by the vowel and vomit clause patterns, the parameters of the vowel and monosyllabic patterns are processed for the uttered syllable (consonant + vowel) pattern. To register as a standard pattern, several types of standard patterns are generated for the single syllable. For example, for the utterance /ma/ of Ma, five vowels a, i.

/a/-/rna/, /i/- with the addition of tsu, 工, and o.
/ma%/u/-/ma/, /e/-/ma/, 10/
-/m&/ standard patterns are created by cutting. The pattern processing at the time of is based on the following formula.

However, Q is "+" - Parameter P is input parameter Wij is the processing weighting coefficient (based on statistical information).Also, it is based on the results of a post-processing device that performs kana-kanji conversion on a phrase-by-phrase basis from the results recognized by the Tsubo syllable pattern. The following equation is used when learning is performed after dividing the input voice in pseudo-continuous utterance into N parts to obtain learning pattern sections.

Q J(t) =+ (,”Ll,P 1(t)Wt
j (t) + Q j (t) ) j = 1 type among 1 to 5 Wij (t) - Wij (Qj (t) / Pz (t)) /
N However, the correspondence between N and the weight time of the number of learning is based on the MDP method.

The figure is a block diagram of a circuit showing one embodiment of the present invention.

First, a single syllable is input, parameter calculation section 1 extracts parameters for pattern matching, and switching section 2
is connected to the monosyllabic standard pattern registration section 10, and the parameters are stored in the monosyllabic standard pattern storage section 11. The monosyllabic standard pattern registration section 10 controls the vowel and monosyllabic standard pattern processing and creation section 9 to send monosyllabic parameters from the monosyllabic standard pattern storage section 11 and record them in the vowel and shore syllable statistical information section 8 in advance. Standard patterns of vowels and pen syllables are created according to the needs of each vowel from the collected information. The monosyllabic standard pattern is stored in the vowel and monosyllabic standard pattern storage section 7, and is used as a one-vehicle weight pattern in the recognition processing section 3 for pattern matching of input by pseudo continuous utterance.

Next, in order to recognize the pseudo-continuous utterances, the pseudo-continuous utterances are input into the input parameter calculation unit 1, the parameters are extracted, the switch 2 is connected to the recognition processing unit 3, and the parameters are converted into vowel and syllable standard patterns. Pattern matching is performed using the vowel and monosyllable standard patterns from the storage section 7, and the recognition result is sent out via the input cover turn sending section 4. Post-processing device i
At t5, the recognition result is subjected to one antonym-kanji conversion for each phrase, and the result is applied to the learning vowel and monosyllable standard pattern processing unit 6.
Send to. Learning vowel and monosyllable standard pattern processing section 6
The recognition result of the input cover turn sending unit 4 and the post-processing device 5
Learning is performed based on the recognition results.

That is, it is corrected by performing a weighted average of the parameters, or if the difference is large, it is replaced, or it is re-registered and stored in the vowel and monosyllabic standard pattern storage section 7.

The learning results are also sent to the vowel and monosyllable statistical information section 58 for correction of the statistical information.

(f) As described in detail, the present invention uses pre-registered standard patterns with only monosyllables, thereby improving the monophonic syllable, which is insufficient for recognizing input speech based on pseudo-continuous utterances. Because it is possible, the effect is like a dog.

[Brief explanation of the drawing]

The figure shows one embodiment of the present invention and is a block diagram of one circuit. 1 is a parameter calculation unit, 2 is a switching unit, 3 is a recognition processing unit,
4 is an input cover pattern sending section, 5 is a post-processing device, 6 is a learning vowel and monosyllable standard pattern processing section, 7 is a vowel and monosyllable standard pattern storage section, 8 is a vowel and monosyllable statistical information section,
Reference numeral 9 represents a vowel and academic syllable standard pattern processing/creation unit, 10 represents a monosyllabic standard pattern registration unit, and 11 represents a monosyllabic standard pattern storage unit.

Claims (1)

[Claims] In a speech recognition device that allows input using pseudo-continuous speech,
Only monosyllables are used as standard patterns to be registered in advance, several types of standard patterns are generated for each monosyllable, and the standard patterns of the numerology are created based on the statistical information of vowels and monosyllables created in advance. Based on the results of a post-processing device that performs kana-kanji conversion on a clause-by-phrase basis, the speech input in pseudo-continuous utterance is subdivided into learning pattern sections, which are then learned. A standard pattern creation method in pseudo-continuous speech recognition characterized by rearranging vowels and monosyllable standard patterns.