Speech recognition method and apparatus

• US 20050288929A1
• Filed: 06/24/2005
• Published: 12/29/2005
• Est. Priority Date: 06/29/2004
• Status: Active Grant

• Current Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Original Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Inventors: Kuboyama, Hideo, Komori, Yasuhiro, Fukada, Toshiaki
• Granted Patent: US 7,565,290 B2
• Time in Patent Office: Days
• US Class Current: 704/251
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]

Abstract

A speech recognition apparatus includes a word dictionary having recognition target words, a first acoustic model which expresses a reference pattern of a speech unit by one or more states, a second acoustic model which is lower in precision than said first acoustic model, selection means for selecting one of said first acoustic model and said second acoustic model on the basis of a parameter associated with a state of interest, and likelihood calculation means for calculating a likelihood of an acoustic feature parameter with respect to said acoustic model selected by said selection means.

Claims

1. A speech recognition apparatus comprising:

a word dictionary having recognition target words;

a first acoustic model which expresses a reference pattern of a speech unit by one or more states;

a second acoustic model which is lower in precision than said first acoustic model;

selection means for selecting one of said first acoustic model and said second acoustic model on the basis of a parameter associated with a state of interest; and

likelihood calculation means for calculating a likelihood of an acoustic feature parameter with respect to said acoustic model selected by said selection means.

2. The apparatus according to claim 1, wherein the parameter indicates a specific ordinal position which the state of interest occupies in said word dictionary from a word start.

3. The apparatus according to claim 1, wherein

said word dictionary forms a tree structure in which the recognition target words share the predetermined speech unit or branches in a network grammar, and

the parameter represents a branch count near the state of interest.

4. The apparatus according to claim 2, wherein said likelihood calculation means calculates a likelihood with respect to a state within a predetermined number of states from the word start in said word dictionary by using said first acoustic model, and calculates a likelihood with respect to a state exceeding a predetermined number of states by using said second acoustic model.

5. The apparatus according to claim 1, wherein with respect to states for which said selection means has selected said second acoustic model, said likelihood calculation means calculates likelihoods by using said second acoustic model and re-calculates likelihoods by using said first acoustic model with respect to only states which exhibit a likelihood higher than a threshold.

6. The apparatus according to claim 1, wherein with respect to states for which said selection means has selected said second acoustic model, said likelihood calculation means calculates likelihoods by using said second acoustic model and re-calculates a likelihood by using said first acoustic model with respect to only a predetermined number of states of the states which are selected in decreasing order of likelihood.

7. The apparatus according to claim 1, wherein the number of mixed distributions which express said second acoustic model is smaller than the number of mixed distributions which express said first acoustic model.

8. The apparatus according to claim 1, wherein

said first acoustic model is a triphone model, and

said second acoustic model is a monophone model.

9. The apparatus according to claim 1, wherein in calculating a likelihood with respect to a state of interest by using said second acoustic model, if likelihood calculation using said first acoustic model has been performed for a state having the same speech unit alignment as that of the state of interest, said likelihood calculation means reuses a result of the likelihood calculation as a result of likelihood calculation for the state of interest.

10. A speech recognition method of performing speech recognition by using a word dictionary having recognition target words, a first acoustic model which expresses a reference pattern of a speech unit by one or two or more states, and a second acoustic model which is lower in precision than the first acoustic model, comprising the steps of:

selecting one of the first acoustic model and the second acoustic model on the basis of a parameter associated with a state of interest; and

calculating a likelihood of an acoustic feature parameter with respect to the selected acoustic model.

11. A program for implementing a speech recognition method defined in claim 10 by using a computer.

12. A computer-readable storage medium in which a program defined in claim 11 is stored.