MODIFYING METHOD FOR SPEECH MODEL AND MODIFYING MODULE THEREOF

• US 20090132249A1
• Filed: 01/10/2008
• Published: 05/21/2009
• Est. Priority Date: 11/21/2007
• Status: Active Grant

• Current Assignee: Industrial Technology Research Institute
• Original Assignee: Industrial Technology Research Institute
• Inventors: Liao, Yuan-Fu, Tu, Jia-Jang
• Granted Patent: US 8,126,711 B2
• Time in Patent Office: Days
• US Class Current: 704/243
• CPC Class Codes: G10L 15/065 Adaptation

Abstract

A modifying method for a speech model and a modifying module thereof are provided. The modifying method is as follows. First, a correct sequence of a speech is generated according to a correct sequence generating method and the speech model. Next, a candidate sequence generating method is selected from a plurality of candidate sequence generating methods, and a candidate sequence of the speech is generated according to the selected candidate sequence generating method and the speech model. Finally, the speech model is modified according to the correct sequence and the candidate sequence. Therefore, the present invention increases a discrimination of the speech model.

Claims

1. A modifying method for a speech model comprising:

generating a correct sequence of a first speech according to a correct sequence generating method and the speech model;

selecting a first candidate sequence generating method from a plurality of the candidate sequence generating methods according to a selection sequence;

generating a candidate sequence of the first speech according to the first candidate sequence generating method and the speech model; and

modifying the speech model according to the correct sequence and the candidate sequence of the first speech.

2. The modifying method for a speech model as claimed in claim 1 further comprising:

receiving a second speech;

generating a correct sequence of the second speech according to the correct sequence generating method and the speech model;

selecting a second candidate sequence generating method from a plurality of the candidate sequence generating methods according to the selection sequence;

generating a candidate sequence of the second speech according to the second candidate sequence generating method and the speech model; and

modifying the speech model according to the correct sequence and the candidate sequence of the second speech.

3. The modifying method for a speech model as claimed in claim 1, wherein the step of selecting the first candidate sequence generating method from a plurality of the candidate sequence generating methods according to the selection sequence further comprises:

setting a control parameter; and

selecting the first candidate sequence generating method from a plurality of the candidate sequence generating methods according to the control parameter, and modifying the control parameter.

4. The modifying method for a speech model as claimed in claim 1, wherein the step of modifying the speech model according to the correct sequence and the candidate sequence of the first speech further comprises:

calculating a misrecognition measure according to the speech model and the correct sequence and the candidate sequence of the first speech;

inputting the misrecognition measure to a loss function; and

modifying the speech model with a fine-tuning approach for minimizing the loss function.

5. The modifying method for a speech model as claimed in claim 1, wherein the selection sequence is an alternate arrangement of the first candidate sequence generating method and a second candidate sequence generating method.

6. The modifying method for a speech model as claimed in claim 1, wherein the selection sequence is an alternate arrangement of the first candidate sequence generating method, the second candidate sequence generating method and a third candidate sequence generating method.

7. The modifying method for a speech model as claimed in claim 1, wherein the candidate sequence generating methods comprise a candidate sequence generating method liable to generate insertion errors, a candidate sequence generating method liable to generate substitution errors and a candidate sequence generating method liable to generate deletion errors.

8. A modifying module for a speech model comprising:

a speech database, for providing a speech;

a correct sequence generator, coupled to the speech database for generating a correct sequence of the speech according to the speech model;

a candidate sequence module, coupled to the speech database, comprising a plurality of candidate sequence generators, and the candidate sequence module selecting a first candidate sequence generator from the candidate sequence generators according to a selection sequence, wherein the first candidate sequence generator generates a candidate sequence of the speech according to the speech model; and

a speech model modifier, coupled to the correct sequence generator and the candidate sequence module for modifying the speech model according to the correct sequence and the candidate sequence of the speech.

9. The modifying module for a speech model as claimed in claim 8, wherein the candidate sequence generators comprise the first candidate sequence generator, a second candidate sequence generator and a third candidate sequence generator, wherein the first candidate sequence generator is liable to generate a candidate sequence with an insertion error, the second candidate sequence generator is liable to generate a candidate sequence with a substitution error, and the third candidate sequence generator is liable to generate a candidate sequence with a deletion error.

10. The modifying module for a speech model as claimed in claim 8, wherein the candidate sequence generators comprise the first candidate sequence generator and a second candidate sequence generator, wherein the selection sequence is an alternate arrangement of the first candidate sequence generator and the second candidate sequence generator.

11. The modifying module for a speech model as claimed in claim 8, wherein the candidate sequence generators comprise the first candidate sequence generator, a second candidate sequence generator and a third candidate sequence generator, and the selection sequence is an alternate arrangement of the first candidate sequence generator, the second candidate sequence generator and the third candidate sequence generator.