JP2966002B2 - Voice recognition device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus for recognizing word speech and sentence speech, and more particularly to an improvement of a speech recognition apparatus for recognizing a large vocabulary.

[Conventional technology]

FIG. 3 shows a speech shown in, for example, a patent application (Japanese Patent Application No. 63-246367) or a collection of lectures by the Acoustical Society of Japan, "Large Vocabulary Word Speech Recognition Using a Linear Context Model" (issued March 14, 1989) 1 is a configuration block diagram illustrating a speech recognition device for large vocabulary configured based on a recognition technology.

In the figure, reference numeral 1 denotes a voice analysis unit for analyzing an input voice;
Generating a model of a speech represented by the symbol series from a sequence of symbols describing the utterance content of the speech depending on the context of the symbol or the context of an input speech portion corresponding to the symbol; Unit 3, a word dictionary describing words to be recognized, 4 a likelihood calculation unit for calculating the likelihood for the input speech in the speech model generated by the model generation unit 2, and 5 a speech model with a high likelihood. This is a primary selection unit as primary selection means for outputting a selected symbol sequence (phoneme sequence) as a source of the generation as a candidate for input speech.

Next, the operation will be described. The input voice is converted into a feature vector sequence by the voice analysis unit 1. For each phoneme sequence of the recognition target word described in the word dictionary 3, the model generation unit 2 extracts context information from the feature vector sequence and the phoneme sequence, and generates a word model in which context is considered based on the context information. Generate. The likelihood calculation unit 4 calculates the likelihood for the feature vector sequence for each word model generated in this manner. The primary selection unit 5 selects N word models with high likelihood among them, and outputs phoneme sequences of these word models as word candidates.

[Problems to be solved by the invention]

Since the conventional speech recognition apparatus configured as described above generates a model in consideration of the context and uses it for recognition, the recognition rate is extremely high, but recognition errors still exist. According to the analysis of the experimental results, most of these recognition errors are errors in similar words in which one to several phonemes in the phoneme sequence are different from the correct phoneme sequence. By the way, conventional devices that consider such contexts have sufficiently higher recognition performance than conventional devices that do not consider context, so even if recognition is incorrect, the correct word is always included in the top word candidates. Be considered
In addition, the first candidate word is likely to be a similar word as described above with respect to the correct word. Therefore, only when words that are similar to the first candidate are included in the top candidates, re-recognition between these similar words using a recognition means suitable for recognizing similar words still results in a higher recognition rate. Should be able to be improved.

However, the above-described conventional apparatus has a problem that even when a similar word is included in the top candidate, a recognition unit suitable for recognizing a similar word is not applied.

The present invention has been made in order to solve the above problems, and recognizes a large vocabulary that includes many similar words in which some symbols (phonemes) in a symbol sequence (phoneme sequence) are different. It is an object of the present invention to provide a speech recognition device with high recognition accuracy even as a target.

[Means for solving the problem]

The speech recognition apparatus according to the present invention includes a similar symbol sequence detection unit (similar phoneme sequence detection) for detecting a similar symbol sequence (phoneme sequence) among the input speech candidates selected by the primary selection unit (primary selection unit 5). Part 6) and an emphasis for generating a speech model in which, when a similar symbol sequence (phoneme sequence) is detected by the similar symbol sequence detecting means, a portion of a different symbol between these similar symbol sequences is emphasized. A model generation unit (emphasis model generation unit 7), and the likelihood calculation unit calculates the likelihood of each emphasized speech model generated by the emphasis model generation unit; Reselecting means (reselection unit 8) for selecting an emphasized speech model having a high likelihood with respect to the input speech from among the emphasized speech models and outputting a series of symbols as a source of the generation as candidates for the input speech.
It is characterized by having.

[Action]

The similar symbol sequence detecting unit (similar phoneme sequence detecting unit 6) detects a similar symbol sequence (phoneme sequence) among the input speech candidates selected by the primary selecting unit (primary selecting unit 5). When a similar symbol sequence (phoneme sequence) is detected by the similar symbol sequence detection unit (similar phoneme sequence detection unit 6), the emphasis model generation unit (emphasis model generation unit 7) A speech model in which different symbol (phoneme) portions are emphasized is generated. The reselecting means (reselecting unit 8) selects an emphasized speech model having a high likelihood for the input speech from among the emphasized speech models generated by the emphasized model generating means (emphasized model generating unit 7), The symbol sequence (phoneme sequence) is output as an input speech candidate.

(Example of the invention)

FIG. 1 is a block diagram showing the configuration of a speech recognition apparatus according to an embodiment of the present invention. 1, components corresponding to those shown in FIG. 3 are denoted by the same reference numerals, and description thereof will be omitted. In FIG. 1, reference numeral 6 denotes a similar phoneme sequence detecting unit as similar symbol sequence detecting means for detecting a similar phoneme sequence (symbol sequence) in the input speech candidates selected by the primary selecting unit 5, and 7 denotes a similar phoneme. When a similar phoneme sequence is detected by the sequence detection unit 6, an emphasis model generation unit as an emphasis model generation unit that generates a speech model emphasizing a different phoneme portion between these similar phoneme sequences, 8 Means for selecting an emphasized speech model having a high likelihood with respect to the input speech from among the emphasized speech models generated by the emphasis model generation unit 7 and outputting a phoneme sequence as a source of the generation as an input speech candidate As the reselection unit.

 Next, the operation will be described.

The input voice is converted into a feature vector sequence by the voice analysis unit 1. On the other hand, the model generation unit 2 extracts the context information from the feature vector sequence and the phoneme sequence of the word dictionary 3, and generates a model considering the context based on the context information. The likelihood calculation unit 4 calculates the likelihood for the input feature vector sequence for each generated model. The primary selection unit 5 selects N models having a high likelihood among them, and sets the phoneme sequence that has been the source of generation of these models as a candidate for the recognition result.

The similar phoneme sequence detection unit 6 detects a different part of the phoneme sequence for a combination of the first-order phoneme sequence and any one of the remaining (N-1) candidates.
FIG. 2 shows that the similar phoneme sequence
6 is a flowchart showing a process for detecting a phoneme, which is different from FIGS. This process includes the following three steps.

[1] Check for dropout (step S1) The sequence A 'formed by removing an arbitrary phoneme (referred to as aj) of the sequence A and the sequence B are checked, and if they match, the phoneme aj of the sequence A is different and is different from the phoneme. I do.

[2] Insertion Check (Step S2) The phoneme sequence B 'formed by removing an arbitrary phoneme (bj) of the sequence B is checked for consistency with the sequence A.
The phoneme bj of the sequence B is different and is a phoneme.

[3] Replacement Check (Step S3) A phoneme sequence A ″ that can be obtained by replacing an arbitrary one phoneme of the sequence A (referred to as aj) with a phoneme at the same position of the sequence B (referred to as bj).
And the sequence B, and if they match, the phoneme of the sequence A
aj and the phoneme bj of the sequence B are different and are phonemes.

In this embodiment, even before re-recognition, the recognition rate is high, and it is rare that two or more phonemes are incorrect.

When generating a model from a certain phoneme sequence C having a different phoneme Q, the emphasis model generation unit 7 generates an emphasis model as follows. First, a normal model for the phoneme sequence C is generated. This normal model consists of a phoneme pattern mean a (s, j) and its covariance matrix C (s, j), a duration mean μ (s, j) and its variance σ ² (s, j). Are connected according to a phoneme sequence based on context information. Here, j is a number representing the position of the phoneme, and s is a number assigned to the internal state of the phoneme. Next, assuming that the position of the phoneme Q is j, the covariance matrix Σ (s, j) (s = 1,2) of the normal model generated in this way
Is replaced by a value obtained by multiplying each element by 1 / w, and further, a variance σ ^{2 of the} duration is replaced by 1 / w to obtain an enhanced model. Here, w must be selected to be larger than 1, and it was best when w = 1.25 experimentally.

The reselection unit 8 calculates the likelihood of the emphasized model for the input speech in the likelihood calculation unit 4 as in the case of the normal model. Here, the likelihood of the emphasized model is different from the likelihood of the normal model, and the covariance of the phoneme pattern and the variance of the phoneme duration of the phoneme part are multiplied by 1 / w. In contrast, since the calculation is performed by being enlarged w times in the different phoneme portions, the different phoneme portions are evaluated with relatively large importance. Therefore, similar words can be identified with high accuracy by performing reselection using the likelihood of the emphasized model.

 The final reselection result is obtained as follows.

[1] A candidate having a phoneme sequence similar to the phoneme sequence of the first candidate (different by one phoneme in the present embodiment) is extracted from the N candidates, and the number thereof is set to M.

[2] For all combinations that can select any two candidates from M candidates having a phoneme sequence similar to the first-order phoneme sequence, as a sequence A and a sequence B, respectively, The likelihood (La) and the likelihood (Lb) of the series B enhancement model are calculated.

[3] Similar M candidates are rearranged according to the likelihood of the emphasized model calculated in [2]. That is, the rearrangement is performed so that the likelihood of the k-th candidate after the rearrangement by the enhancement model is higher than the likelihood of the k + 1-th and lower candidates by the enhancement model.

[4] A model having a high likelihood is selected as a recognition result by the rearrangement in [3].

In the above description, the symbol describing the utterance of the voice is a phoneme, but may be another symbol such as a phoneme or a phoneme fragment. In addition, although words are targeted for recognition, not only words but also phrases and morphemes can be targeted for recognition. further,
As similar words, words that differ by one phoneme in the phoneme sequence are detected, but words that differ by two or more phonemes may be detected as similar words.

〔The invention's effect〕

As described above, according to the present invention, a similar symbol sequence detecting unit that detects a similar symbol sequence among the input speech candidates selected by the primary selecting unit, and a similar symbol sequence detected by the similar symbol sequence detecting unit. An emphasis model generation means for generating a speech model emphasizing a different symbol portion between these similar symbol sequences when detected, a likelihood of each emphasis speech model generated by the emphasis model generation means Is calculated by the likelihood calculating means, and an emphasized speech model having a high likelihood with respect to the input speech is selected from among the emphasized speech models generated by the emphasized model generating means, and a symbol as a source of the generation is selected. And a re-selecting unit that outputs the sequence as a candidate of the input speech. Therefore, when a similar word is detected in the top candidate of the recognition result, a different part of the symbol sequence between the detected similar words is used. Note Is generated, a model having a high likelihood for the input speech is selected from the emphasized models, and a result of re-recognition is obtained, whereby similar words in which some symbols in the symbol sequence are different from each other are obtained. There is an effect that a high recognition accuracy can be obtained even when a large vocabulary including many vocabularies is targeted for recognition.

[Brief description of the drawings]

FIG. 1 is a block diagram showing the configuration of a speech recognition apparatus according to an embodiment of the present invention, FIG. 2 is a flowchart showing the operation of a similar phoneme sequence detection unit in this embodiment, and FIG. It is a block diagram. 2... Model generation unit (model generation unit), 5... Primary selection unit (primary selection unit), 6... Similar phoneme sequence detection unit (similar symbol sequence detection unit), 7. Generating means), 8... Reselection unit (reselection means).

──────────────────────────────────────────────────続 き Continued on front page (58) Field surveyed (Int.Cl. ⁶ , DB name) G10L 3/00-9/20

Claims (1)

(57) [Claims] 1. Model generation means for generating a speech model represented by a symbol sequence from a sequence of symbols describing the utterance content of the speech depending on the context of the symbol or the context of an input speech portion corresponding to the symbol. And likelihood calculating means for calculating the likelihood of each of the voice models generated by the model generating means with respect to the input voice; and a voice having a high likelihood for the input voice among the voice models generated by the model generating means. A primary selecting means for selecting a model and outputting a series of symbols as a source of the generation as a candidate for an input voice, wherein a similarity among the input voice candidates selected by the primary selecting means is provided. A similar symbol sequence detecting means for detecting a symbol sequence of the symbol sequence, and a sound emphasizing a different symbol portion between these similar symbol sequences when the similar symbol sequence is detected by the similar symbol sequence detecting device. Emphasis model generation means for generating a voice model, and the likelihood calculation means for calculating the likelihood of each emphasized speech model generated by the emphasis model generation means; and Reselecting means for selecting an emphasized speech model having a high likelihood with respect to an input speech from among the emphasized speech models and outputting a series of symbols as a source of the generation as a candidate for the input speech. Recognition device.