Speech recognition using middle-to-middle context hidden markov models

• US 5,812,974 A
• Filed: 04/10/1996
• Issued: 09/22/1998
• Est. Priority Date: 03/26/1993
• Status: Expired due to Term

• Current Assignee: Texas Instruments, Inc.
• Original Assignee: Texas Instruments, Inc.
• Inventors: Hemphill, Charles T., Kribs, Christopher M., Netsch, Lorin P.
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): Wieland, Susan
• Application Number:
  US08/631,731
• Time in Patent Office: 895 Days
• Field of Search: 395/2.5, 395/2.64, 395/2.65, 395/2.66, 704/241, 704/255, 704/256, 704/257
• US Class Current: 704/256.4
• CPC Class Codes: G10L 15/18 using natural language modelling
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 2015/088 Word spotting
• CPC Class Codes: G10L 25/87 Detection of discrete points within a voice signal

Abstract

This is a speech recognition method for modeling adjacent word context, comprising: dividing a first word or period of silence into two portions; dividing a second word or period of silence, adjacent to the first word, into two potions; and combining last portion of the first word or period of silence and first portion of the second word or period of silence to make an acoustic model. The method includes constructing a grammar to restrict the acoustic models to the middle-to-middle context.

Claims

1. A speech recognition method for modeling adjacent word context, comprising:

a. dividing a first word or period of silence into two portions;

b. dividing a second word or period of silence, adjacent to said first word or period of silence, into two portions; and

c. combining last portion of said first word or period of silence and first portion of said second word or period of silence to make an acoustic model.

2. The method of claim 1, wherein said portions constitute about half of each word, and wherein each word is divided in a stable acoustic context.

3. The method of claim 1, wherein said acoustic model is restricted to the appropriate context by constructing grammar rules.

4. The method of claim 1, wherein said grammar comprises:

utilizing at least three grammar rules;

wherein a first grammar rule starts with a silence model;

wherein a last grammar rule ends with a silence model; and

wherein middle grammar rule contains nonterminal symbols that constrain a second part of an acoustic model, representing the first half of a word or period of silence, to match the first part of an adjacent acoustic model, representing the second half of said word or period of silence.

5. The method of claim 1, wherein said acoustic models are created utilizing Hidden Markov Modeling techniques.

6. The method of claim 1, wherein said acoustic models are created utilizing neural network modeling techniques.

7. The method of claim 1, wherein said acoustic models are created utilizing Dynamic Time Warping Template modeling techniques.

8. A speech recognition system utilizing acoustic models for modeling adjacent word context, comprising:

a. a speech recognizing device;

b. a means for separating words and periods of silence, wherein said means for separating words is connected to said speech recognizing device;

c. a memory for storing separated words and periods of silence;

d. a computational means for creating an acoustic model of said words and periods of silence;

e. a means for dividing said acoustic models of said words and periods of silence into two portions;

f. a means for combining last portion of a first word and first portion of a second word or period of silence to result in a new acoustic model.

9. The system of claim 8, wherein said system includes grammar rules coupled to computational means for combining acoustic models.

10. The system of claim 9, wherein said acoustic models are restricted to the appropriate context by said grammar rules.

11. The system of claim 10, wherein said grammar rules comprises:

at least three grammar rules;

wherein a first grammar rule starts with a silence model;

a last grammar rule ends with a silence model; and

wherein middle grammar rule contains nonterminal symbols that constrain a second part of an acoustic model, representing the first half of a word or period of silence, to match the first part of an adjacent acoustic model, representing the second half of said word or period of silence.

12. The system of claim 8, wherein said system utilizes Hidden Markov Models.

13. The system of claim 8, wherein said system utilizes neural network acoustic models.

14. The system of claim 8, wherein said system utilizes Dynamic Time Warping Template models.

15. A speech recognition method for modeling adjacent word context, comprising:

a. dividing a first word or period of silence into two portions;

b. dividing a second word or period of silence, adjacent to said first word or period of silence, into two portions;

said portions constitute about half of each word, and wherein each word is divided in a stable acoustic context; and

c. combining last portion of said first word or period of silence and first portion of said second word or period of silence to make an acoustic model;

said acoustic model is restricted to the appropriate context by constructing grammar rules;

said grammar comprises;

utilizing at least three grammar rules;

wherein a first grammar rule starts with a silence model;

wherein a last grammar rule ends with a silence model; and

wherein middle grammar rule contains nonterminal symbols that constrain a second part of an acoustic model, representing the first half of a word or period of silence, to match the first part of an adjacent acoustic model, representing the second half of said word or period of silence.

16. A speech recognition system utilizing acoustic models for modeling adjacent word context, comprising:

a. a speech recognizing device;

b. a means for separating words and periods of silence, wherein said means for separating words is connected to said speech recognizing device;

c. a memory for storing separated words and periods of silence;

d. a computational means for creating an acoustic model of said words and periods of silence;

e. a means for dividing said acoustic models of said words and periods of silence into two portions;

f. a means for combining last portion of a first word and first portion of a second word or period of silence to result in a new acoustic model;

said system includes grammar rules coupled to computational means for combining acoustic models;

said acoustic models are restricted to the appropriate context by said grammar rules;

wherein said grammar rules comprises;

at least three grammar rules;

wherein a first grammar rule starts with a silence model;

a last grammar rule ends with a silence model; and

wherein middle grammar rule contains nonterminal symbols that constrain a second part of an acoustic model, representing the first half of a word or period of silence, to match the first part of an adjacent acoustic model, representing the second half of said word or period of silence.