Technique for training a phonetic decision tree with limited phonetic exceptional terms

• US 8,027,834 B2
• Filed: 06/25/2007
• Issued: 09/27/2011
• Est. Priority Date: 06/25/2007
• Status: Active Grant

• Current Assignee: Cerence Operating Company (Cerence Inc.)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Hancock, Steven M.
• Primary Examiner(s): Azad; Abul K
• Application Number:
  US11/767,751
• Publication Number:
  US 20080319753A1
• Time in Patent Office: 1,555 Days
• Field of Search: 704230-269
• US Class Current: 704/254
• CPC Class Codes: G10L 13/04 Details of speech synthesis systems, e.g. synthesiser structure or memory management

Abstract

The present invention discloses a method for training an exception-limited phonetic decision tree. An initial subset of data can be selected and used for creating an initial phonetic decision tree. Additional terms can then be incorporated into the subset. The enlarged subset can be used to evaluate the phonetic decision tree with the results being categorized as either correctly or incorrectly phonetized. An exception-limited phonetic tree can be generated from the set of correctly phonetized terms. If the termination conditions for the method have been determined to be unsatisfactorily met, then steps of the method can be repeated.

Claims

1. A method for testing a phonetic decision tree, the method comprising:

testing a first phonetic decision tree, created using a first set of data, with a second set of data, the second set of data including at least one second term not in the first set of data, by phonetizing terms of the second set of data using the first phonetic decision tree;

categorizing results of the testing into a set of correctly phonetized terms and a set of incorrectly phonetized terms; and

operating at least one processor to create an exception dictionary including at least one term from the set of incorrectly phonetized terms and phonetization information related to the at least one term.

2. The method of claim 1, wherein the categorizing further comprises:

generating a phonetization corresponding to a term in the second set of data;

comparing the generated phonetization to a standard pronunciation of the term;

when the generated phonetization is equivalent to the standard pronunciation, classifying the term as correctly phonetized; and

when the generated phonetization does not match the standard pronunciation, classifying the term as incorrectly phonetized.

3. The method of claim 2, wherein the generating the phonetization, comparing the generated phonetization, and classifying the term are repeated for each term contained within the first set of data.

4. The method of claim 2, wherein generating the phonetization comprises generating a speech output, and

wherein comparing the generated phonetization to the standard pronunciation of the term comprises comparing the speech output to the standard pronunciation.

5. The method of claim 1, wherein the testing, categorizing, and creating are performed in a development environment of a text-to-speech (TTS) system.

6. The method of claim 1, further comprising:

using the exception dictionary by a speech synthesis engine in a runtime environment of a TTS system to produce a speech output corresponding to a term in the set of incorrectly phonetized terms.

7. The method of claim 1, wherein the testing, categorizing, and creating are performed by at least one machine in accordance with at least one computer program stored in a computer readable storage media, said computer programming having a plurality of code sections that are executable by the at least one machine.

8. The method of claim 1, wherein the accepting comprises:

following the categorizing, determining if one or more termination conditions are met by the first phonetic decision tree; and

when the one or more termination conditions are met, accepting the first phonetic decision tree as a production decision tree to be used in the text-to-speech system.

9. The method of claim 8, further comprising:

when the one or more termination conditions are not met, repeating, until the one or more termination conditions are met, acts of

creating a new phonetic decision tree using a new set of data including data not previously used to create a phonetic decision tree,

following the expanding, testing the new phonetic decision tree with at least one third term not in the new set of data by phonetizing the at least one third term using the new phonetic decision tree,

categorizing results of the testing into a set of correctly phonetized terms and a set of incorrectly phonetized terms, and

determining if the one or more termination conditions are met by the new phonetic decision tree; and

when the one or more termination conditions are met, accepting the new phonetic decision tree as the production decision tree to be used in the text-to-speech system.

10. The method of claim 9, further comprising:

following each categorizing during the repeating, removing from the new set of data terms categorized into the set of incorrectly phonetized terms, and

wherein categorizing a result into the set of incorrectly categorized terms comprises adding the term to the set of incorrectly categorized terms.

11. The method of claim 9, further comprising, during the repeating:

prior to each creating of the new decision tree during the repeating, removing terms from the set of incorrectly categorized terms.

12. The method of claim 11, wherein creating the exception dictionary from the set of incorrectly phonetized terms comprises creating the exception dictionary comprising at least one term from the set of incorrectly phonetized terms from a last categorizing prior to determining that the one or more termination conditions are met.

13. A system for phonetic decision tree testing, the system comprising:

at least one recording device to store processor-executable instructions; and

at least one processor coupled to the at least one recording device and programmed by the processor-executable instructions to act as;

a training engine configured to evaluate a phonetic decision tree, created using a first data set of terms of a training data set, using a second set of terms from the training data set and a set of standard pronunciations, the second set of terms comprising terms for evaluating the phonetic decision tree and including at least one second term not in the first data set, wherein the training engine categorizes a result of the evaluation into a set of correctly phonetized terms and a set of incorrectly phonetized terms,

wherein the phonetic tree generation engine is configured to create an exception-limited phonetic decision tree from the set of correctly phonetized terms.

14. The system of claim 13, wherein the training engine further comprises:

a training interface configured to provide user-configuration of the training data set and one or more termination conditions.

15. The system of claim 13, wherein the phonetic tree generation engine is further configured to create an exception dictionary for use in the text to speech processing, the exception dictionary comprising at least one term of the set of incorrectly phonetized terms and pronunciation information related to the at least one term.

16. The system of claim 13, wherein the phonetic tree generation engine is configured to create an exception-limited phonetic decision tree by:

following the evaluation and the categorizing by the training engine, determining if one or more termination conditions are met by the phonetic decision tree; and

when the one or more termination conditions are met, adopting the phonetic decision tree as the exception-limited phonetic decision tree.

17. The system of claim 16, wherein the phonetic tree generation engine is configured to, when the one or more termination conditions are not met by the phonetic decision tree, repeat, until the one or more termination conditions are met:

creating a new phonetic decision tree using a new data set including the first data set and an additional data set; and

based on evaluation and categorizing of the new phonetic decision tree by the training engine, determine whether the one or more termination conditions are met by the new phonetic decision tree; and

when the one or more termination conditions are met by the new phonetic decision tree, adopt the new phonetic decision tree as the exception-limited phonetic decision tree.

23. The method of claim 18, wherein the acts of claim 15 are performed automatically by at least one machine in accordance with at least one computer program having a plurality of code sections that are executable by the at least one machine, said at least one computer program being stored in a machine readable storage medium.

18. A method for creating a phonetic tree for speech synthesis comprising acts of:

generating an initial phonetic tree from a first training data set of terms and corresponding term pronunciations;

creating a second data set including at least one second term not in the first training data set;

creating at least one phonetization for at least one term in the second data set using the initial phonetic tree;

comparing a first phonetization of the at least one phonetization for a first term of the at least one term to a corresponding correct term pronunciation for the first term;

categorizing at least one result of the comparing into a set of correctly phonetized terms and a set of incorrectly phonetized terms; and

creating an exception dictionary including at least one incorrectly-phonetized term of the set of incorrectly phonetized terms and phonetization information related to the at least one incorrectly-phonetized term.

19. The method of claim 18, further comprising:

determining whether a termination condition has been reached based at least in part on the categorizing, wherein the termination condition is based at least in part upon a number of terms categorized into the set of incorrectly phonetized terms;

when the termination condition has been reached, accepting the phonetic tree as a production tree to be used by a speech synthesis engine to generate speech output from text input; and

when the termination condition has not been reached, repeating the generating, creating, creating, comparing to generate and evaluate a different phonetic tree until the termination condition is reached, and accepting the different phonetic tree as the production tree when the termination condition is created.

20. The method of claim 18, further comprising:

establishing a frequency list of words in a language sorted by frequency of use, and creating the training set from N percentage of words in the frequency list, wherein N is a configurable percentage.

21. The method of claim 18, further comprising:

using the exception dictionary of words at runtime, by the speech synthesis engine, in conjunction with a phonetic decision tree.

22. The method of claim 21, further comprising:

removing from the expanded data set those words that were not correctly text-to-speech converted using the initial phonetic tree;

wherein creating the exception dictionary comprises utilizing a set of words removed from the expanded data set by the removing step.