Speech recognition apparatus and speech recognition method

• US 7,310,601 B2
• Filed: 12/08/2005
• Issued: 12/18/2007
• Est. Priority Date: 06/08/2004
• Status: Active Grant

• Current Assignee: Panasonic Intellectual Property Corporation of America (Panasonic Holdings Corporation)
• Original Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Inventors: Nishizaki, Makoto, Nakatoh, Yoshihisa, Yamada, Maki, Yoshizawa, Shinichi
• Primary Examiner(s): Lerner, Martin
• Application Number:
  US11/296,268
• Publication Number:
  US 20060100876A1
• Time in Patent Office: 740 Days
• Field of Search: 704/270, 704/236, 704/240, 704/231, 704/255, 704/250, 704/239, 704/257
• US Class Current: 704/240
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/32 Multiple recognisers used in sequence or in parallel; Score combination systems therefor, e.g. voting systems

Abstract

The present invention provides a speech recognition apparatus which appropriately performs speech recognition by generating, in real time, language models adapted to a new topic even in the case where topics are changed. The speech recognition apparatus includes: a word specification unit for obtaining and specifying a word; a language model information storage unit for storing language models for recognizing speech and corresponding pieces of tag information for each language model; a combination coefficient calculation unit for calculating the weights of the respective language models, as combination coefficients, according to the word obtained by the word specification unit, based on a relevance degree between the word obtained by the word specification unit and the tag information of each language model; a language probability calculation unit for calculating the probabilities of word appearance by combining the respective language models according to the calculated combination coefficients; and a speech recognition unit for recognizing speech by using the calculated probabilities of word appearance.

Claims

1. A speech recognition apparatus which obtains and recognizes speech, said apparatus comprising:

a language model storage unit operable to store language models for recognizing speech;

a word obtainment unit operable to obtain one of a plurality of words;

a specific probability derivation unit operable to derive, for each of the language models, a specific model probability which is a probability that a predetermined word appears next to the word obtained by said word obtainment unit, based on the language model;

a tag information storage unit operable to store a piece of tag information for each of the language models, the tag information indicating a feature of each language model;

a relevance degree holding unit operable to hold a relevance degree between each piece of tag information and each of the words;

an importance degree holding unit operable to hold an importance degree of each piece of tag information to a corresponding one of the language models;

a relevance degree derivation unit operable to derive the relevance degree between each piece of tag information and the word obtained by said word obtainment unit, from the respective relevance degrees held by said relevance degree holding unit;

a combination coefficient calculation unit operable to calculate, as a combination coefficient, a weight of each language model which corresponds to the obtained word, based on relevance degrees derived by said relevance degree derivation unit between each piece of tag information and the obtained word, and the importance degrees held by said importance degree holding unit, each of the relevance degrees indicating a relevance degree between the obtained word and one of the pieces of tag information of each language model;

a probability calculation unit operable to calculate a probability that the predetermined word appears next to the obtained word by calculating, for each of the language models, a product of the specific model probability and the combination coefficient, and by adding the obtained products of the respective language models, the specific model probability corresponding to the language model derived by said specific probability derivation unit, and the combination coefficient corresponding to the language model calculated by said combination coefficient calculation unit; and

a recognition unit operable to recognize the speech by using the probability calculated by said probability calculation unit,

wherein said word obtainment unit is operable to obtain the one of the words adapted to the speech recognized by said recognition unit.

2. The speech recognition apparatus according to claim 1, wherein said combination coefficient calculation unit is operable to calculate a combination coefficient of each language model, each time a word is obtained by said word obtainment unit.

3. The speech recognition apparatus according to claim 1, wherein said combination coefficient calculation unit is operable to calculate a combination coefficient of each language model, each time plural words are obtained by said word obtainment unit.

4. The speech recognition apparatus according to claim 1, wherein said combination coefficient calculation unit is operable to calculate a weight, as a combination coefficient, of each language model corresponding to plural words obtained by said word obtainment unit, based on the relevance degree between the plural words obtained by said word obtainment unit and the tag information of each language model.

5. The speech recognition apparatus according to claim 1, further comprising a keyword extraction unit operable to extract a keyword from at least one of electronic data that a user is browsing and profile information related to the user,

wherein said word obtainment unit is operable to obtain, as the obtained one of the words, the keyword extracted by said keyword extraction unit.

6. A speech recognition method for obtaining speech and recognizing data stored in a recording medium,

wherein the recording medium includes:

a language model storage unit operable to store language models for recognizing speech;

a tag information storage unit operable to store a piece of tag information for each of the language models, the tag information indicating a feature of each language model;

a relevance degree holding unit operable to hold a relevance degree between each piece of tag information and each of a plurality of words; and

an importance degree holding unit operable to hold an importance degree of each piece of tag information to a corresponding one of the language models;

wherein said speech recognition method comprises;

obtaining one of the words;

deriving, for each of the language models, a specific model probability which is a probability that a predetermined word appears next to the word obtained in said obtaining, based on the language model;

deriving the relevance degree between each piece of tag information and the word obtained by said obtaining of the word, from the respective relevance degrees held by the relevance degree holding unit;

calculating, as a combination coefficient, a weight of each language model which corresponds to the obtained word, based on relevance degrees derived by said deriving of the relevance degree between each piece of tag information and the obtained word, and the importance degrees held by the importance degree holding unit, each of the relevance degrees indicating a relevance degree between the obtained word and one of the pieces of tag information of each language model;

calculating a probability that the predetermined word appears next to the obtained word by calculating, for each of the language models, a product of the specific model probability and the combination coefficient, and by adding the obtained products of the respective language models, the specific model probability corresponding to the language model derived in said deriving of the specific model probability, and the combination coefficient corresponding to the language model calculated in said calculating of the weight as the combination coefficient; and

recognizing the speech by using the probability calculated in said calculating of the probability,

wherein said obtaining of the one of the words includes obtaining the one of the words adapted to the speech recognized in said recognizing of the speech.

7. A program, recorded on a computer-readable recording medium, for causing a computer to obtain speech and recognize the speech by using data stored on the recording medium,

wherein the recording medium includes:

a language model storage unit operable to store language models for recognizing speech;

a tag information storage unit operable to store a piece of tag information for each of the language models, the tag information indicating a feature of each language model;

a relevance degree holding unit operable to hold a relevance degree between each piece of tag information and each of a plurality of words; and

an importance degree holding unit operable to hold an importance degree of each piece of tag information to a corresponding one of the language models;

wherein said program is for causing a computer to execute operations comprising;

obtaining one of the words;

deriving, for each of the language models, a specific model probability which is a probability that a predetermined word appears next to the word obtained in said obtaining, based on the language model;

deriving the relevance degree between each piece of tag information and the word obtained by said obtaining of the word, from the respective relevance degrees held by the relevance degree holding unit;

calculating, as a combination coefficient, a weight of each language model which corresponds to the obtained word, based on relevance degrees derived by said derivation of the relevance degree between each piece of tag information and the obtained word, and the importance degrees held by the importance degree holding unit, each of the relevance degrees indicating a relevance degree between the obtained word and one of the pieces of tag information of each language model;

calculating a probability that the predetermined word appears next to the obtained word by calculating, for each of the language models, a product of the specific model probability and the combination coefficient, and by adding the obtained products of the respective language models, the specific model probability corresponding to the language model derived in said deriving of the specific model probability, and the combination coefficient corresponding to the language model calculated in said calculating of the weight as the combination coefficient; and

recognizing the speech by using the probability calculated in said calculating of the probability; and

wherein said obtaining of the one of the words includes obtaining the one of the words adapted to the speech recognized in said recognizing of the speech.