Method of calculating HMM output probability and speech recognition apparatus

• US 20030061044A1
• Filed: 07/18/2002
• Published: 03/27/2003
• Est. Priority Date: 07/24/2001
• Status: Active Grant

• Current Assignee: Seiko Epson Corporation (Seiko Group)
• Original Assignee: Seiko Epson Corporation (Seiko Group)
• Inventors: Miyazawa, Yasunaga, Hasegawa, Hiroshi
• Granted Patent: US 7,058,576 B2
• Time in Patent Office: Days
• US Class Current: 704/256
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]

Abstract

In speech recognition based on HMM, in which speech recognition is performed by performing vector quantization and obtaining an output probability by table reference, the amount of computation and use of memory area are minimized while achieving a high ability of recognition.

Codebooks used for vector quantization are, for example, if phonemes are used as subwords, codebooks for respective phonemes, such that a codebook CB1 is a codebook for a phoneme /a/ and a codebook CB2 is a codebook for a phoneme /i/, and these codebooks are associated with respective phoneme HMMs 11 and 12. When a feature vector obtained by speech analysis is vector quantized based on, for example, the codebook CB1 and a code (label) is output, tables T1, T2, and T3 for respective states of the phoneme HMM 11 associated with the codebook CB1 are each referred to in order to obtain state output probabilities corresponding to the label, and speech recognition is performed using the state output probabilities as a parameter.

Claims

1. ) a method of calculating an hmm output probability, in which a feature vector obtained by speech analysis of an input speech is vector quantized using vector quantization means and a state output probability in accordance with a result of quantization is obtained by referring to tables,

wherein the vector quantization means used for vector quantization are vector quantization means associated with respective subword, and the tables are provided in association with the vector quantization means associated with the subwords, and

the feature vector obtained by speech analysis of the input speech is vector quantized using one of the vector quantization means associated with subwords, a result of quantization is output, and a state output probability in accordance with the result of quantization is obtained by referring to a table associated with the vector quantization means that has output the result of quantization:

2. ) A method of calculating an HMM output probability according to claim 1, wherein the subwords are phonemes, vector quantization means associated with phonemes are created, and the vector quantization means associated with phonemes are associated with respective phoneme HMMs.

3. ) A method of calculating an HMM output probability according to claim 1, wherein the subwords are syllables, vector quantization means associated with syllables are created, and the vector quantization means associated with syllables are associated with respective syllable HMMs.

4. ) A method of calculating an HMM output probability according to claim 1, wherein the subwords are vowels and consonants, and the vector quantization means associated with subwords are vector quantization means created from vowels and vector quantization means created from consonants.

5. ) A method of calculating an HMM output probability according to one of claims 1 to 3, wherein, of the vector quantization means associated with subwords, vector quantization means that can be shared, if any, are shared on the basis of each group of the respective subword HMMs.

6. ) A method of calculating an HMM output probability according to one of claims 1 to 3, wherein, of the vector quantization means associated with subwords, vector quantization means that can be shared, if any, are shared on the basis of each state of the respective subword HMMs.

7. ) A method of calculating an HMM output probability according to one of claims 1 to 6, wherein each of the vector quantization means associated with subwords has an optimal code size in accordance with the associated subword.

8. ) A method of calculating an HMM output probability according to one of claims 1 to 7, wherein whether an input subword is a subword to be vector quantized is checked before performing vector quantization, and vector quantization is skipped if the input subword is not a subword to be vector quantized.

9. ) A method of calculating an HMM output probability according to one of claims 1 to 8, wherein the vector quantization means are codebooks.

10. ) A speech recognition apparatus for vector quantizing a feature vector obtained by speech analysis of an input speech using vector quantization means, obtaining a state output probability in accordance with a result of quantization by referring to tables, and performing speech recognition using the state output probability as a parameter, said speech recognition apparatus comprising:

a speech analysis unit for performing speech feature analysis of an input speech to output a feature vector sequence;

a vector quantization means storage unit storing vector quantization means associated with respective subwords;

a vector quantization processing unit for vector quantizing the feature vector sequence output from the speech analysis unit using the vector quantization means associated with subwords, stored in the vector quantization storage means, and outputting a result of quantization;

a subword HMM storage unit including various parameters required for subword HMM calculation and tables for referring to state output probabilities, provided in association with the vector quantization means associated with subwords; and

a speech recognition processing unit for referring to the table associated with vector quantization means used for vector quantization upon receiving a result of quantization from the vector quantization means, obtaining a state output probability in accordance with the result of quantization, and performing speech recognition using the state output probability as a parameter to output a result of recognition.

11. ) A speech recognition apparatus according to claim 10, wherein the subwords are phonemes, vector quantization means associated with phonemes are created, and the vector quantization means associated with phonemes are associated with respective phoneme HMMs.

12. ) A speech recognition apparatus according to claim 10, wherein the subwords are syllables, vector quantization means associated with syllables are created, and the vector quantization means associated with syllables are associated with respective syllable HMMs.

13. ) A speech recognition apparatus according to claim 10, wherein the subwords are vowels and consonants, and the vector quantization means associated with subwords are vector quantization means created from vowels and vector quantization means created from consonants.

14. ) A speech recognition apparatus according to one of claims 10 to 12, wherein, of the vector quantization means associated with subwords, vector quantization means that can be shared, if any, are shared on the basis of each group of the respective subword HMMs.

15. ) A speech recognition apparatus according to one of claims 10 to 12, wherein, of the vector quantization means associated with subwords, vector quantization means that can be shared, if any, are shared on the basis of each state of the respective subword HMMs.

16. ) A speech recognition apparatus according to one of claims 10 to 15, wherein each of the vector quantization means associated with subwords has an optimal code size in accordance with the associated subword.

17. ) A speech recognition apparatus according to one of claims 10 to 16, wherein a word/language table is provided, whether an input subword is a subword to be vector quantized is checked by referring to the word/language table before the vector quantization unit performs vector quantization, and vector quantization is skipped if the input subword is not a subword to be vector quantized.

18. ) A speech recognition apparatus according to one of claims 10 to 17, wherein the vector quantization means are codebooks.