JP2012255867A - Voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a robust voice recognition device having a lower false recognition rate.SOLUTION: The voice recognition device includes: a sound feature amount converting part 12 having a language model 22 that puts a limit on a phoneme sequence pattern, a word phoneme label dictionary 21 storing a phoneme label, a label conversion rule dictionary 23 storing a rule for converting the phoneme label, and an acoustic model 24 for generating a standard voice pattern to convert the voice to a feature amount; a phoneme label converting part 13 that refers to the language model 22, the word phoneme label dictionary 21, and the label conversion rule dictionary 23, to convert the voice converted by the sound feature amount converting part to a phoneme label; a similarity calculation part 14 that converts the phoneme label converted by the phoneme label converting part 13 to a standard voice pattern by the acoustic model 24 to calculate similarity to the voice converted by the sound feature amount converting part 12; and a maximum likelihood grammar deciding part 15 that determines an input sentence from a result obtained by the similarity calculation part 14.

Description

  The present invention relates to a speech recognition apparatus.

  In recent years, command type speech recognition systems have been used. In the command-type speech recognition system, the recognition grammar can be fixed in advance because the tasks are limited. In speech recognition control, the recognition grammar is used as a constraint condition for phoneme sequences (acoustic model chain patterns).

Conventionally, a method of assigning a normal reading kana and a phoneme pattern on a one-to-one basis has been used, and this method is effective in many cases in speech recognition. However, even if speech recognition is performed using this method, the recognition error rate may be 5% to 50%.
In such a case, commands that are easily misrecognized are biased, and the results of misrecognition are often biased. That is, there are patterns that are easy to make mistakes.

  Patent Document 1 discloses a speech recognition method when the S / N ratio is reduced due to the presence of noise. According to this, acoustic fluctuations can be absorbed by preparing acoustic models having different S / N ratios in advance. In addition, a phoneme pattern is generated in advance within a range in which a phoneme can be converted by noise.

  Patent Document 2 discloses a correction location determination apparatus that corrects a misrecognition portion of a speech recognition location correctly and efficiently. According to this, when the character string portion whose pronunciation is similar to the correct character string does not exist in the speech recognition result, the corrected part determining means inserts the correct character string using the language model describing the word connection constraint. A position with a high probability is searched as an insertion position, and the recognition result correcting means inserts a correct character string at the determined insertion position.

  Patent Document 3 discloses a speech recognition apparatus that obtains a speech recognition result with high robustness while improving semantic suitability. According to this, N or more maximum likelihood solutions using the speech feature amount extracted by the micro extraction unit, the phoneme HMM stored in the speech model storage unit, and the word 2-gram model stored in the language model storage unit And its likelihood (score) is calculated.

JP-A-8-123467 JP 2006-267319 A JP 2005-211752 A

[Search on May 10, 2011], Internet <URL: http://shower.human.waseda.ac.jp/~m-kouki/pukiwiki_public/66.html> [Search May 10, 2011], Internet <URL: http://www.slp.is.ritsumei.ac.jp/C/pattern-rec/pr.pdf>

  In the case of a method in which a normal reading kana and a phoneme pattern are assigned one-to-one, erroneous recognition is likely to occur when a phoneme disappears, when there is a phoneme conversion, and when there is a phoneme intermediate. When the phoneme disappears, for example, “i” is not pronounced for the pronunciation of “Akita” but becomes “a k (i) t a”. The case where there is a phoneme conversion is a case where “book” is “h o N” but “one” is “i q p o N”. For example, there is a phoneme neutralization, for example, the right roll is pronounced “migir o_ ru” but the phoneme “ir” is replaced by “y” and becomes “migy o_ ru”. is there.

  In such a case, the recognition result is affected by the phoneme pattern that appears first, and erroneous recognition is likely to occur. For example, if the recognition grammar contains “Akita”, it may be easily mistaken for “Akita”, and even if the user utterance is “Akita”, the recognition result may be “芥”. Similarly, if there is “right yaw”, it is easily mistaken for “right roll”, and even if the user utterance is “right roll”, the recognition result may be “right yaw”.

The speech recognition apparatus comprises: a language model that stores a fixed grammar that restricts a phoneme sequence pattern; a word phoneme label dictionary that stores a phoneme label that divides a sentence extracted by the language model into phonemes; and the word phoneme When a phoneme pattern divided by the label dictionary has a specific pattern, the phoneme pattern has a label conversion rule dictionary that stores rules for converting a phoneme model, and an acoustic model that models a standard phoneme pattern. A phoneme label conversion unit that converts a sound signal into a phoneme label with reference to an acoustic feature value conversion unit that converts a speech signal into a feature value, the language model, the word phoneme label dictionary, and the label conversion rule dictionary, and the phoneme label The phoneme label converted by the conversion unit is converted into a voice pattern based on the acoustic model, and the acoustic feature value conversion unit converts the phoneme label into a feature value. Comprises a similarity calculating unit calculating a similarity between the audio signal, based on the calculation result of the similarity calculation unit, and a maximum likelihood grammar determining section determines an appropriate input sentence.
As a result, phoneme labels can be converted online for a specific pattern without the need for prior learning of an acoustic model or a language model.

  Provided is a speech recognition device that is highly robust and has a reduced misrecognition rate.

It is a figure which shows the structural article of the speech recognition apparatus concerning Embodiment 1. FIG. It is a figure which shows the phoneme label concerning Embodiment 1. FIG. It is a figure which shows the structural article and data of the speech recognition apparatus concerning Embodiment 1. FIG.

Embodiment 1
Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing the components of the voice recognition device 1. The speech recognition apparatus 1 includes a speech input unit 11, an acoustic feature amount conversion unit 12, a phoneme label conversion unit 13, a similarity calculation unit 14, a maximum likelihood grammar determination unit 15, a word phoneme label dictionary 21, a language model 22, and a label conversion rule. A dictionary 23 and an acoustic model 24 are included.

  The voice input unit 11 collects voice uttered by the user. For example, the voice input unit 11 is a microphone. The voice input unit 11 outputs the input voice to the acoustic feature amount conversion unit 12.

  The acoustic feature amount conversion unit 12 samples the collected voice and performs voice analysis on the sampled voice data. As a result, the acoustic feature value conversion unit 12 extracts feature values for each fixed section. For example, the acoustic feature value conversion unit 12 performs voice feature value conversion using MFCC (Mel Frequency Cepstral Coefficient) (Non-Patent Document 1). The acoustic feature amount conversion unit 12 outputs the voice that has been converted into the feature amount to the similarity calculation unit 14.

  The phoneme label conversion unit 13 converts the phoneme label by referring to the word phoneme label dictionary 21, the language model 22, and the label conversion rule dictionary 23, respectively. The phoneme label conversion unit 13 outputs the stored phoneme label to the similarity calculation unit 14.

  The word phoneme label dictionary 21 stores labels provided for each phoneme. FIG. 2 is an example of a phoneme label stored in the word phoneme label dictionary 21. When a description is made using a word phoneme label, a long sound is described without a space. For example, “a_” for “A” and “yo_ka N” for “Yokan”.

The language model 22 is a fixed grammar that fixedly defines a morpheme sequence and restricts a phoneme sequence pattern. That is, the language model 22 is a model of word chains based on grammar and statistics.
For example, in the language model 22, if the sound data input to the phoneme label conversion unit 13 is “Tell me your phone number”, “Sato” “san” “no” “phone” “number” “ For each element of "Teach" and "Te", the appearance pattern is restricted based on the grammar and statistical model. More specifically, it is limited that words such as “san” and “you” appear after “Sato”. For example, after “san”, it is limited that words such as “no” and “ha” appear. Similarly, the appearance pattern is restricted for other words. The language model 22 can perform the same processing using an N-gram with probability.

In the label conversion rule dictionary 23, if a specific pattern exists in a phoneme label defined by the combination of the word phoneme label dictionary 21 and the language model 22, a rule for conversion to a corresponding label pattern is recorded. For example, if there is a phoneme pattern {ir}, the rule is to change it to {iyr}. Typically, the label conversion rule dictionary 23 can arbitrarily add or delete rules.
If the phoneme pattern conversion unit 13 includes a phoneme pattern recorded in the label conversion rule dictionary 23, the phoneme label conversion unit 13 records the phoneme pattern changed according to the rule. Note that it is desirable that the phoneme label conversion unit 13 records both phoneme models before and after the phoneme pattern change.

The similarity calculation unit 14 refers to the acoustic model 24 and calculates the similarity between the standard phoneme pattern and the input speech pattern under linguistic restrictions. More specifically, the similarity calculation unit 14 generates a speech pattern generated by referring to the acoustic model 24 for the phoneme label input from the phoneme label conversion unit 13, and a speech input from the acoustic feature amount conversion unit 12. Calculate the similarity to the pattern.
The similarity calculation unit 14 outputs the calculated similarity to the maximum likelihood grammar determination unit 15.

In the acoustic model 24, a standard pattern of a combination of one phoneme and other phonemes before and after the phoneme is recorded. For example, the phoneme {m} is recorded with a combination pattern with the phonemes {a, i, etc.} before and after the phoneme {m}. Furthermore, the acoustic model 24 records pronunciations according to phoneme combinations.
The similarity calculation unit 14 creates a standard speech pattern based on the phoneme label input from the phoneme label conversion unit 13 and the acoustic model 24, and the featured speech that is the output of the acoustic feature amount conversion unit 12. Calculate similarity.

  The maximum likelihood grammar determination unit 15 determines the grammar having the highest similarity based on the similarity calculated by the similarity calculation unit 14.

  Next, the operation of the voice recognition device 1 will be described. In the following description, it is assumed that “right roll” is input to the voice input unit 11 as voice. FIG. 3 is a diagram showing the components of the voice recognition device 1 and data when the “right roll” is voice-input to the voice input unit 11.

  The voice input unit 11 collects voice uttered by the user. The voice input unit 11 outputs the collected voice to the acoustic feature amount conversion unit 12.

  The acoustic feature quantity conversion unit 12 analyzes the voice signal input to the voice input unit 11, cuts out a voice section delimited by silence, and converts it into a feature quantity. The acoustic feature amount conversion unit 12 outputs the voice pattern that is converted into the feature amount to the similarity calculation unit 14.

The phoneme label conversion unit 13 converts phoneme labels based on the rules.
Specifically, the phoneme label conversion unit 13 refers to the language model 22. As a result, the phoneme label conversion unit 13 uses the grammar and statistical model for the appearance pattern of each element of “right”, “roll”, and “direction” in the sentence “right roll direction” input to the speech input unit 11. Give limits based on. More specifically, the phoneme label conversion unit 13 refers to the language model 22 to extract “roll” and “yaw” connected to the word “right” as a grammar.
Next, the phoneme label conversion unit 13 refers to the word phoneme label dictionary 21 for the words extracted using the language model 22, and “right” {migi}, “roll” {ro_ru}, “yaw”. {Yo_} is extracted. That is, the phoneme label conversion unit 13 stores {migir o_ ru} as “right roll” and {migiy o_} as “right yaw”.

Next, the phoneme label conversion unit 13 refers to the label conversion rule dictionary 23. Here, it is assumed that the rule for converting {ir} to {iyr} is described in the label conversion rule dictionary 23.
When the “right roll direction” is described by a phoneme label, {migir o_ ruho o_ k o_} and there is a rule for converting {ir} to {iyr} in the label conversion rule dictionary 23. The unit 13 converts {migir o_ruh o_ k o_} into {migiyr o_ ruh o_ k o_} and stores it. Note that the phoneme label conversion unit 13 stores both the phoneme model before conversion based on the label conversion rule dictionary 23 and the phoneme model after conversion. That is, the phoneme label conversion unit 13 records both {migir o_ ru} and {migiyr o_ ru} as the “right roll”.
Note that the label conversion rule dictionary 23 does not record a rule for converting a phoneme label included in “right yaw”. Therefore, the phoneme label conversion unit 13 is in a state where {migiy o_} is still recorded as “right yaw”.

  The similarity calculation unit 14 applies the standard phoneme pattern recorded in the acoustic model 24 for each of {migir o_ ru}, {migiyr o_ru}, and {migiy o_} recorded in the phoneme label conversion unit 13. I do. Thereby, the similarity calculation unit 14 generates a speech pattern for each of {m i g i ro_r u}, {m i g i y r o_ ru}, and {m i g i y o_}.

  The similarity calculation unit 14 calculates the similarity between a plurality of speech patterns generated using the phone model labels recorded in the acoustic model 24 and the phoneme label conversion unit 13 and the speech pattern input from the acoustic feature amount conversion unit 12. calculate. That is, with respect to the voice pattern input from the acoustic feature amount conversion unit 12, the similarity of the voice pattern of {migiro_ru} and {migiyr_o} as “right roll” and {migiy_o_} as “right yaw” is set. calculate. Typically, the similarity calculation unit 14 calculates the similarity between the frequency of the voice of the generated voice pattern and the frequency of the voice of the voice pattern input from the acoustic feature amount conversion unit 12 (Non-Patent Document 2). .

  The maximum likelihood grammar determination unit 15 determines the highest similarity among the similarities calculated by the similarity calculation unit 14. For example, the similarity calculation unit 14 determines that the “right roll” has been input if the calculation result indicates that the similarity of {m i g i y r o — ru u} of the language model is the highest.

Thereby, the speech recognition apparatus 1 can reduce the misrecognition rate during speech recognition.
The speech recognition apparatus 1 can change a specific phoneme pattern to a corresponding phoneme pattern. This change target is not an acoustic model but a language model. Therefore, when a change is made in the acoustic model, learning by the speech recognition apparatus 1 and data collection are required. However, the user can arbitrarily change the change pattern of the label conversion rule dictionary 23 without performing learning or data collection. Can be set.

  Note that the present invention is not limited to the above-described embodiment, and can be changed as appropriate without departing from the spirit of the present invention. For example, the label conversion rule dictionary 23 has been described on the assumption that a rule for converting {ir} to {iyr} is described, but a larger number of conversion rules are stored and used for conversion of phoneme labels. Also good. Furthermore, when a plurality of conversion rules registered in the label conversion rule dictionary 23 correspond to one phoneme model, the phoneme label conversion unit 13 applies only the first conversion rule, for example. The phoneme labels after conversion may be generated in various combinations, such as a case where only the second conversion rule is applied, and a case where both the first and second rules are applied.

DESCRIPTION OF SYMBOLS 1 Speech recognition apparatus 11 Speech input part 12 Acoustic feature-value conversion part 13 Phoneme label conversion part 14 Similarity calculation part 15 Maximum likelihood grammar determination part 21 Word phoneme label dictionary 22 Language model 23 Label conversion rule dictionary 24 Acoustic model

Claims (1)

A language model that stores fixed grammars that restrict phoneme sequence patterns;
A word phoneme label dictionary that stores phoneme labels that divide sentences extracted by the language model into phonemes;
A label conversion rule dictionary for storing rules for converting a phoneme model when there is a specific pattern in the phoneme pattern divided by the word phoneme label dictionary;
An acoustic model storing a standard phoneme pattern;
An acoustic feature amount conversion unit for converting the input audio signal into a feature amount;
A phoneme label conversion unit for converting to a phoneme label with reference to the language model, the word phoneme label dictionary, and the label conversion rule dictionary;
A similarity calculation unit that converts the phoneme label converted by the phoneme label conversion unit into a speech pattern based on the acoustic model, and calculates a similarity with the speech signal characterized by the acoustic feature amount conversion unit; ,
A speech recognition apparatus comprising: a maximum likelihood grammar determination unit that determines an appropriate input sentence based on a calculation result by the similarity calculation unit.

Images