JP2951332B2 - Clause candidate reduction method in speech recognition - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Overview] A phrase candidate reduction method in phrase speech recognition in which effective candidates to be recognized are selected and reduced in advance when performing phrase vocabulary recognition of a large vocabulary. The purpose of the present invention is to provide a phrase candidate reduction scheme in speech recognition that extracts features and reduces the amount of processing in a phrase created by combining words. For each word in the word dictionary containing all the words to be recognized, a feature phoneme that is expected to appear when the phrase containing the word is uttered is extracted by the word feature phoneme extraction unit, and the word feature phoneme is extracted. The phrase feature phonemes of less than the number of words extracted by the extraction unit are combined in the phrase feature phoneme synthesis unit to synthesize the features as phrases, and input. The speech is analyzed by the analysis unit, and the feature phonemes of the input speech are extracted from the output by the input phoneme extraction unit, based on the syllable features synthesized by the syllable feature phoneme synthesis unit and the phonemes of the input speech extracted by the input phoneme extraction unit. Thus, a phrase candidate selection unit selects a phrase candidate effective for the collation processing of speech recognition.

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a phrase candidate reduction method in phrase speech recognition that selects and reduces valid candidates to be recognized in advance when performing phrase speech recognition of a large vocabulary.

2. Description of the Related Art In recent years, a speech recognition device has been used for various purposes, and is useful when, for example, a document is created by speech input, or an instruction to the device is given by speech. In such a case, it is desired that the speech is not simply recognized in word units but is recognized by a phrase which is a longer unit, and it is also required to recognize a phrase voice having a large vocabulary. However, for that purpose, the phrase to be recognized becomes enormous, and the time required for the recognition process becomes longer, and its improvement is desired.

[Prior Art] FIG. 3 shows a configuration diagram of a conventional example.

The configuration shown in FIG. 3 is a word candidate reduction device in speech recognition developed by the same applicant as the present application (Japanese Patent Application No. Sho 62-53066).
No.) is applied to phrase speech recognition.

In the figure, the word dictionary 30 stores all words constituting a phrase to be recognized together with its phoneme label network. The phoneme label network associates phoneme labels indicating each consonant or vowel so as to express the voice of a word, and an example is shown in FIG. Figure 4 shows the Japanese word (also a phrase) "but"
In the figure, "#" is a word boundary, "SH" is a consonant part of "shi",
“I” indicates the unvoiced vowel “A”, “i” indicates the unvoiced vowel “A”, “K” indicates the consonant part of “K”, and “A” indicates the vowel “A”. It is a phoneme label. In this manner, a voice deformation phenomenon called voicelessness is expressed using a network expression.

The phrase synthesizing unit 31 synthesizes all possible phrases from the word dictionary 30.

The phrase characteristic phoneme extraction unit 32 extracts a characteristic phoneme for each synthesized phrase. That is, a characteristic phoneme that is expected to appear more reliably for each phrase is extracted in advance.

On the other hand, the input voice is subjected to short-time frequency analysis by the analysis unit 35 to generate short-time spectrum time-series data (referred to as a spectrum pattern).
Characteristic phonemes are detected from 35 output data. As the characteristic phoneme, a phoneme pronounced with a large power is easily and reliably detected, and is extracted as a characteristic phoneme and supplied to the phrase candidate selection unit 33.

The phrase candidate selection unit 33 calculates a correlation between each characteristic phoneme extracted by the input characteristic phoneme extraction unit 34 and the characteristic phoneme of each phrase obtained from the phrase feature phoneme extraction unit 32, and performs speech recognition. Select a phrase that is valid for the matching process. In the process of the phrase candidate selection unit 33, a phrase in which the characteristic phoneme of each phrase is similar to the characteristic phoneme extracted from the input speech is selected as a phrase to be a candidate when performing phrase speech recognition collation.
Note that the matching operation is to calculate the similarity (or distance) between the spectrum pattern of the input speech and the spectrum pattern synthesized from the phoneme label network as a candidate, and to detect and output the most similar phrase. DP for correcting the time axis deviation between the input pattern and the phrase that is the matching candidate
A matching method using a dynamic programming matching method is used as a high-performance processing method.

[Problems to be Solved by the Invention] According to the configuration of the conventional example described above, all possible phrases are synthesized from the word dictionary, so that the number of phrases is enormous, and the processing time accompanying the processing amount is enormous. There was a problem of becoming.

An object of the present invention is to provide a phrase candidate reduction method in speech recognition that extracts features from words in a word dictionary and reduces the amount of processing in a phrase created by combining words.

[Means for Solving the Problems] FIG. 1 shows a principle configuration diagram of the present invention.

In FIG. 1, 10 is a word dictionary, 11 is a word feature phoneme extraction unit that extracts feature phonemes of each word, 12 is a phrase feature phoneme synthesis unit that combines words and combines each feature phoneme, and 13 is used for matching. A phrase candidate selection unit that selects phrase candidates, 14 is an input phoneme extraction unit that extracts phonemes from an output obtained by analyzing input speech, and 15 represents an analysis unit.

The present invention extracts a characteristic phoneme of a word in a word dictionary,
A syllable that is a feature of a syllable composed of words is synthesized by a combination of feature syllables of the word and input to the syllable candidate selection unit. Is selected as a candidate for collation.

[Operation] Each word in the word dictionary 10 of FIG. 1 stores its part of speech and its inflected form, and it is expected that the word feature phoneme extraction unit 11 will surely appear when a phrase including the word is uttered in advance. Characteristic phonemes are extracted.

When the characteristic phoneme is extracted for each word, the phrase characteristic phoneme synthesis unit 13 extracts a characteristic phoneme when a phrase is synthesized by combining words according to a rule determined for each word. This extraction is performed by combining characteristic phonemes of words.

On the other hand, the analysis unit 15 analyzes the input voice in the same manner as in the related art, and extracts phonemic candidates of the input voice from the analysis output.

The phonemes (including candidates) extracted by the input phoneme extraction unit 14 and the feature phonemes of many phrases synthesized by the phrase feature phoneme synthesis unit 12 are supplied to the phrase candidate selection unit 13, and all the feature phonemes are included in the input phoneme candidates. Is selected as a candidate phrase used for matching, and a phrase including a feature phoneme that is not included is not adopted as a candidate for matching.

[Embodiment] FIG. 2 shows a configuration diagram of an embodiment of the present invention.

In FIG. 2, 20 is a word dictionary, 21 is a feature phoneme rule storage unit, 22 is a word feature phoneme extraction unit, 23 is a word feature phoneme pattern holding unit, 24 is a phrase model storage unit, 25 is a phrase feature phoneme synthesis unit, 26 is a phrase feature phoneme pattern holding unit, 27 is a selection unit, 28 is a phrase candidate generation unit, 29 is an input phoneme extraction unit, and 40 is an analysis unit.

In FIG. 2, the word dictionary 20 contains all words to be recognized. Each word describes the part of speech and inflected form, and also describes the phoneme label network of the word.

The word characteristic phoneme extraction unit 22 extracts a characteristic phoneme that is expected to appear surely when a phrase including the word is uttered, for each word in the word dictionary according to the characteristic phoneme rule. Therefore, for example, the inflected ending does not always appear, and it is difficult to extract a characteristic phoneme therefrom. For example, in the case of "ARUKU", "KU" changes (like "does not exist") due to utilization,
Characteristic phonemes are extracted from “ARU”.

Also, for example, for the word "SHIKASHI (but)",
This is not a characteristic phoneme because I may be unvoiced and not uttered, and the characteristic phoneme pattern of this word is "SH,
A, SH ".

The characteristic phoneme rule storage unit 21 stores rules for extracting characteristic phonemes. (Rules such that a phoneme that is not vocalized and not uttered does not become a feature phoneme).

Thus, the characteristic phoneme pattern of each word is extracted and stored in the word characteristic phoneme pattern holding unit 23.

Next, the phrase feature phoneme synthesis unit 25 outputs a phrase model storage unit 24.
According to the phrase model described in (1), a phrase characteristic phoneme is synthesized by combining a characteristic phoneme pattern of another word that can be combined with a characteristic phoneme pattern of the word based on the part of speech information of each word. Clause model storage unit 24
Are, for example, "noun + particle", "verb + auxiliary verb + particle", and the like. The synthesizing of the phrase feature phoneme synthesis unit 25 is performed by connecting the word feature phoneme pattern holding units 23 to be combined in series.

The output synthesized by the phrase feature phoneme synthesis unit 25 is stored in the phrase feature phoneme pattern holding unit 26.

The input speech is subjected to short-time frequency analysis in the analysis unit 40, and the result is supplied to the input phoneme extraction unit 29, and the input phoneme candidate is extracted. Here, a strong phoneme is detected based on the analysis data, and a phoneme pattern is created. For example, if you say “but,” you ’ll get “(SH, S), I, (K,
T, P), A, (SH, S) ". Here, parentheses mean that either of them is taken.

The selection unit 27 checks whether each pattern in the phrase feature phoneme pattern holding unit 26 is included in all the phoneme candidates of the input speech supplied from the input phoneme extraction unit 29, and includes all the patterns. Select only the phrase feature phoneme patterns that are present. For example, if the input voice is “(S
H, S), I, (K, T, P), A, (SH, S) ", among the phrase feature phoneme patterns," SH, A, SH "and" S, A, SH " Selected.

The phrase candidate generation unit 28 detects the original word feature phoneme pattern that synthesizes the phrase feature phoneme pattern selected by the selection unit 27 by referring to the clause feature phoneme pattern holding unit 26, and further detects the word feature phoneme pattern. The original word from which the pattern was obtained is detected by referring to the word characteristic phoneme pattern holding unit 23. A phrase candidate is generated by synthesizing the detected words.

For example, if the selected phrase feature phoneme pattern is “SH, A, S
If the word is "H", the only word feature pattern that synthesizes it is "SH, A, SH", and the word from which the word feature phoneme pattern can be obtained is "SHIKASHI (but)" or "SHIT
ASHII. Furthermore, phrases are synthesized therefrom, and "SHIKASHI", "SHTASHII (close)", "SH
"ITASHIKU" is a candidate for a phrase.

It goes without saying that the information of each phrase that is a phrase candidate is collated by the collation unit of the recognition device with the analysis data of the input voice by the conventional technique, and recognition is performed.

[Effects of the Invention] According to the present invention, features are extracted at the word level without synthesizing all possible phrases as in the prior art, and selecting the phrase candidates. Is synthesized, the number of phrases to be selected can be reduced, and the process of candidate selection can be speeded up.

[Brief description of the drawings]

FIG. 1 is a block diagram of the principle of the present invention, FIG. 2 is a block diagram of an embodiment of the present invention, FIG. 3 is a block diagram of a conventional example, and FIG. 4 is an explanatory diagram of a phoneme label network. In FIG. 1, 10: word dictionary 11: word feature phoneme extraction unit 12: phrase feature phoneme synthesis unit 13: phrase candidate selection unit 14: input phoneme extraction unit 15: analysis unit

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-62-32499 (JP, A) JP-A-63-85697 (JP, A) JP-A-61-238099 (JP, A) JP-A-60-1985 33599 (JP, A) JP-A-63-220298 (JP, A) JP-A-58-87599 (JP, A) JP-A-61-256396 (JP, A) JP-A-61-62167 (JP, A) JP-B 5-67040 (JP, B2) (58) Fields investigated (Int. Cl. ⁶ , DB name) G10L 3/00-9/20

Claims (1)

(57) [Claims] In a phrase candidate reduction method in a speech recognition device using phrases as a recognition unit, for each word in a word dictionary including all words to be recognized, the phrase appears when a phrase including the word is uttered. In the word feature phoneme extraction unit, the feature phonemes that are reliably predicted are extracted by the word feature phoneme extraction unit. The input speech is analyzed by the analysis unit, the characteristic phonemes of the input speech are extracted from the output by the input phoneme extraction unit, and the syllable features synthesized by the phrase feature phoneme synthesis unit and the input speech extracted by the input phoneme extraction unit are synthesized. A phrase candidate reduction method in speech recognition, characterized in that a phrase candidate selection unit selects a phrase candidate effective for a speech recognition collation process based on a phoneme of the phrase.