Method of phonetic modeling using acoustic decision tree

• US 6,317,712 B1
• Filed: 01/21/1999
• Issued: 11/13/2001
• Est. Priority Date: 02/03/1998
• Status: Expired due to Term

• Current Assignee: Texas Instruments, Inc.
• Original Assignee: Texas Instruments, Inc.
• Inventors: Kondo, Kazuhiro, Kao, Yu-Hung
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US09/235,031
• Time in Patent Office: 1,027 Days
• Field of Search: 704/256, 704/257, 704/251, 704/254, 704/255, 704/209
• US Class Current: 704/256.3
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 2015/022 Demisyllables, biphones or triphones being the recognition units
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering

Abstract

Phonetic modeling includes the steps of forming triphone grammars (11) from phonetic data, training triphone models (13), clustering triphones (14) that are acoustically close together and mapping unclustered triphone grammars into a clustered model (16). The clustering process includes using a decision tree based on the acoustic likelihood and allows sub-model clusters in user-definable units.

Claims

1. A method of performing speech recognition comprising the steps of:

generating phonetic models comprising the steps of forming triphone grammars from phonetic data;

training triphone models;

clustering triphones that are acoustically close together to form clustered triphone model by an acoustic decision tree analysis; and

mapping unclustered triphone grammars into a clustered model; and

recognizing input speech by comparing said input speech to said clustered triphone model.

2. The method of claim 1 wherein said clustering step has flexible submodel grouping of cluster sizes.

3. The method of claim 2 wherein said sub-model grouping is based on the class of phone in which said grouping reside.

4. The method of claim 1 wherein said clustering is a likelihood improvement criterion.

5. The method of claim 1 wherein said clustering triphones clusters said triphones such that the cluster size per phone is based on the entropy of the phone.

6. The method of claim 1 wherein the clustering triphones includes division of the phone class into clusters based on acoustic likelihood.

7. The method of claim 1 wherein the clustering is based on the weight of the acoustic likelihood calculation by the entropy of the cluster in question.

8. The method of claim 7 wherein said decision tree analysis includes decision criteria based on regular expression as a pattern match.

9. A speech recognition system comprising:

a microphone for receiving speech;

a clustered model from clustering triphones that are acoustically close together; and

a processor including a comparison means coupled to said microphone and said clustered model and responsive to said speech received for comparing incoming speech to said clustered model to provide a given output when there is a compare.

10. The recognition system of claim 9 wherein said clustering triphones that are acoustically close together is by clustering by an acoustic decision tree analysis.

11. A speech recognition system comprising:

a clustered model from clustering triphones according to the steps of;

collecting speech data, forming triphone grammars, clustering triphones that are acoustically close together and clustering triphones by decision tree analysis wherein the decision criteria is on likelihood improvement based on acoustic vectors; and

a speech recognizer for comparing said incoming speech to said clustered model for recognizing speech.