Word predicting method, voice recognition method, and voice recognition apparatus and program using the same methods

• US 20030046073A1
• Filed: 08/22/2002
• Published: 03/06/2003
• Est. Priority Date: 08/24/2001
• Status: Active Grant

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: International Business Machines Corporation
• Inventors: Mori, Shinsuke, Nishimura, Masafumi, Itoh, Nobuyasu
• Granted Patent: US 7,480,612 B2
• Time in Patent Office: Days
• US Class Current: 704/251
• CPC Class Codes: G10L 15/193 Formal grammars, e.g. finite state automata, context free grammars or word networks

Abstract

A word predicting method for use with a voice recognition using a computer includes the steps of specifying a sentence structure of a history up to a word immediately before the word to be predicted, referring to a context tree stored in arboreal context tree storage section having information about possible structures of a sentence and a probability of appearance of words with respect to the structures at nodes, and predicting words based on the context tree and the specified sentence structure of the history.

Claims

1. A word predicting method of predicting words in a predetermined sentence by using a computer, comprising the steps of:

acquiring a history to be used in predicting words from a history storage means where a word prediction history is stored and specifying a sentence structure of said history;

acquiring a context tree for word prediction from a context tree storage means, which stores said context tree having information about possible structures of a sentence and a probability of appearance of words with respect to said structures at nodes; and

predicting a word to be predicted based on said specified sentence structure of said history and said acquired context tree for word prediction.

2. The word predicting method according to claim 1, wherein said history is a row of partial parse trees, and the possible structures of the sentence at nodes of said context tree comprise a tree structure, and said word predicting method further comprises a step of predicting a word to be predicted by comparing a tree consisting of a virtual root having said row of partial parse trees directly under it added to said row of partial parse trees with said tree structure at the nodes of said context tree.

3. The word predicting method according to claim 1, further comprising the step of:

acquiring a context tree for sentence structure prediction from said context tree storage unit, which stores the context tree having the information about possible structures of the sentence and the probability of appearance of the sentence structure following said structures at nodes; and

predicting the sentence structure containing a predicted word, based on said predicted word, said sentence structure used in predicting said word, and said acquired context tree for sentence structure prediction, and storing said sentence structure in said history storage unit.

4. A word predicting method of predicting words in a predetermined sentence by using a computer, comprising the steps of:

acquiring a history to be used in predicting words from a history storage unit where a word prediction history is stored and specifying a sentence structure of said history;

deciding an allowance range for predicting a word to be predicted in said sentence structure, based on the specified sentence structure of said history; and

predicting the word to be predicted, based on said sentence structure of said history in said decided allowance range.

5. The word predicting method according to claim 4, wherein said history is a row of partial parse trees.

6. The word predicting method according to claim 4, further comprising the steps of:

deciding an allowance range in said sentence structure for predicting a new sentence structure containing said predicted words, based on said specified sentence structure of said history; and

predicting a new sentence structure containing said predicted words based on said sentence structure of said history in said decided allowance range and said predicted words, and storing it in said history storage unit.

7. A voice recognition method of recognizing a voice signal as a word string by using a computer, comprising the steps of:

making an arithmetical operation on the voice signal to be processed, using an acoustic model and selecting a word as a recognition candidate resulted from the arithmetical operation;

specifying a sentence structure of a history up to the word immediately before the word to be predicted for said selected word as an object; and

predicting said word to be predicted based on a context tree having the information about possible structures of a sentence and a probability of appearance of words with respect to said structures at nodes and the sentence structure of said history.

8. A voice recognition method of recognizing a voice signal as a word string by using a computer, comprising the steps of:

making an arithmetical operation on the voice signal to be processed, using an acoustic model and selecting a word as a recognition candidate resulted from the arithmetical operation;

specifying a sentence structure of a history up to the word immediately before the word to be predicted for the selected word as an object;

deciding an allowance range for predicting the word to be predicted in said sentence structure, based on the specified sentence structure of said history; and

predicting said word to be predicted based on the sentence structure of said history in said decided allowance range.

9. A data processing method comprising the steps of:

acquiring a processing history of a tree structure to be used in predicting a predetermined element from a history storage unit storing said processing history for an array;

acquiring a stochastic model from a stochastic model storage unit storing the stochastic model for the tree structure having predetermined partial trees and a probability distribution associated with said partial trees at nodes; and

retrieving nodes corresponding to the tree structure of said processing history for said stochastic model, and predicting said predetermined element based on said probability distribution associated with said nodes.

10. A voice recognition apparatus, comprising:

an acoustic processing section for converting an analog voice signal into a digital voice signal;

an acoustic model storage section for storing an acoustic model;

a context tree storage section for storing a context tree having the information about possible structures of a sentence and a probability of appearance of words with respect to said structures at nodes; and

a decoding section for recognizing said digital voice signal converted in said acoustic processing section as a word string, using said acoustic model and said context tree,

wherein said decoding section scans said context tree, based on the sentence structure of said history for word prediction, and predicts a word to be predicted, based on the probability of appearance at nodes corresponding to the sentence structure of said history.

11. A voice recognition apparatus, comprising:

an acoustic processing section for converting an analog voice signal into a digital voice signal;

an acoustic model storage section for storing an acoustic model;

a context tree storage section for storing a context tree having the information about possible structures of a sentence and a probability of appearance of words with respect to said structures at nodes; and

a decoding section for recognizing said digital voice signal converted in said acoustic processing section as a word string, using said acoustic model and said context tree,

wherein said decoding section decides an allowance range in said sentence structure for predicting a word to be predicted, based on the sentence structure of a word prediction history, and predicts the word to be predicted, based on the sentence structure of said history in said allowance range.

12. A program for predicting words in a predetermined sentence by controlling a computer,

wherein said program comprises:

a process for acquiring a history to be used in predicting words from history storage means where a word prediction history is stored and specifying a sentence structure of said history;

a process for acquiring a context tree for word prediction from context tree storage means, which stores the context tree having the information about possible structures of a sentence and a probability of appearance of words with respect to said structures at nodes; and

a process for predicting a word to be predicted, based on said specified sentence structure of said history and said acquired context tree for word prediction.

13. A program for predicting words in a predetermined sentence by controlling a computer, wherein said program comprises:

a process for acquiring a history to be used in predicting words from a history storage means where a word prediction history is stored and specifying a sentence structure of said history;

a process for deciding an allowance range in said sentence structure for predicting a word to be predicted, based on the specified sentence structure of said history; and

a process for predicting the word to be predicted, based on the sentence structure of said history in said decided allowance range.

14. A recording medium for storing the data making up a stochastic model, said recording medium being readable on a computer,

wherein said stochastic model has a tree structure;

a root of said stochastic model has a label of empty tree; and

each node of said stochastic model has a label of a partial tree in which a tree as label at a parent node is expanded at predetermined node, in which a probability distribution to be returned in scanning said stochastic model under the condition of the predetermined partial parse tree is appended to the node having the label of corresponding partial tree.

15. The recording medium according to claim 14, wherein the label of each node in said stochastic model is the partial tree decided by applying a certain evaluation function to all the forms in which the tree as label at the parent node is expandable.

16. The recording medium according to claim 14, wherein the label of each node in said stochastic model is the partial tree expanded based on a comparison result that is obtained by making a comparison between the evaluation function when the tree as label at the parent node is expanded at each node and when not expanded.