Granted Patent B1

US 8,065,144 · App. 12/699,172 · Granted Nov 22, 2011

Abstract

Claims (44)

1. A computer-implemented method in which a computer system initiates execution of software instructions stored in memory, the computer-implemented method comprising:

generating a set of estimated pronunciations from text spellings of a set of acoustic training words, each pronunciation comprising a grouping of subword units at least some of which are common to two or more languages, the set of acoustic training words comprising at least a first word and a second word, the first word having a pronunciation based on utterances of native speakers of a first language, the second word having a pronunciation based on utterances of native speakers of a second language;

mapping sequences of sound associated with utterances of each of the acoustic training words against the estimated pronunciation associated with each of the acoustic training words;

using the mapping of sequences of sound to estimated pronunciations to generate a single acoustic subword model for each of the subword units in the grouping of subwords, the acoustic subword model comprising a sound model and a subword unit; and

using the single acoustic subword model for speech recognition.

2. The computer-implemented method of claim 1 further comprising:

generating, for each word in a list of words to be recognized, an acoustic word model, the generating comprising generating a grouping of subword units representing a pronunciation of a respective word to be recognized.

3. The computer-implemented method of claim 2 , wherein the grouping of subwords is a network, and the network represents two pronunciations of a word, the two pronunciations being representative of utterances of native speakers of two languages.

4. The computer-implemented method of claim 2 further comprising:

processing an utterance; and

scoring matches between the processed utterance and the acoustic word models.

5. The computer-implemented method of claim 4 , wherein the utterance is spoken by a native speaker of one of a plurality of languages.

6. The computer-implemented method of claim 4 , wherein the utterance is spoken by a native speaker of a language other than a plurality of languages, the language having similar sounds and similar letters to sounds rules as a language from the plurality of languages.

7. The computer-implemented method of claim 4 , wherein generating the acoustic word model, processing the utterance, and scoring matches is executed by a portable programmable device.

8. The computer-implemented method of claim 2 , wherein the grouping of subword units is a linear sequence of subword units.

9. The computer-implemented method of claim 1 , wherein generating the set of estimated pronunciations includes using a decision tree to map letters of text spellings to pronunciation subword units.

10. The computer-implemented method of claim 1 , wherein using the mapping of sequences of sound to estimated pronunciations to generate the single acoustic subword model for each of the subword units includes mixing distributions of acoustic parameters representing the sounds of the subword unit in multiple languages when a subword unit is common to two or more languages.

11. The computer-implemented method of claim 1 , wherein a first training word in a first set of acoustic training words corresponds to a first language and a second training word in a second set of acoustic training words corresponds to a second language, the first and second acoustic training words having identical text spellings, wherein received pronunciations for the first and second acoustic training words being different.

12. The computer-implemented method of claim 11 , wherein utterances of the first and the second acoustic training words are used to train a common subset of subword units.

13. The computer-implemented method of claim 1 further comprising:

accepting a plurality of sets of utterances, each set corresponding to a different one of the plurality of languages, the utterances in each set being spoken by native speakers of a language of each set; and

training a set of acoustic models for the subword units using the accepted sets of utterances and estimated pronunciations from text representations of the sets of utterances.

14. The computer-implemented method of claim 1 , wherein generating a set of estimated pronunciations further comprises:

forming, from sequences of letters of each acoustic training word's textual spelling and a corresponding grouping of subword units of each respective pronunciation, a letter to subword mapping for each acoustic training word; and

using letter-to-subword mappings to train for generating estimated pronunciations.

15. A computer program product, tangibly embodied in a non-transitory storage medium, the computer program product being operable to cause data processing apparatus to:

generate a set of estimated pronunciations from text spellings of a set of acoustic training words, each pronunciation comprising a grouping of subword units at least some of which are common to two or more languages, the set of acoustic training words comprising at least a first word and a second word, the first word having a pronunciation based on utterances of native speakers of a first language, the second word having a pronunciation based on utterances of native speakers of a second language;

map sequences of sound associated with utterances of each of the acoustic training words against the estimated pronunciation associated with each of the acoustic training words;

use the mapping of sequences of sound to estimated pronunciations to generate a single acoustic subword model for each of the subword units in the grouping of subwords, the acoustic subword model comprising a sound model and a subword unit; and

use the single acoustic subword model for speech recognition.

16. The computer program product of claim 15 , the computer program product being further operable to cause the data processing apparatus to:

generate, for each word in a list of words to be recognized, an acoustic word model, the generating comprising generating a grouping of subword units representing a pronunciation of a respective word to be recognized.

17. The computer program product of claim 16 , wherein the grouping of subwords is a network, and the network represents two pronunciations of a word, the two pronunciations being representative of utterances of native speakers of two languages.

18. The computer program product of claim 15 , the computer program product being further operable to cause the data processing apparatus to:

process an utterance; and

score matches between the processed utterance and the acoustic word models.

19. The computer program product of claim 18 , wherein the utterance is spoken by a native speaker of a language other than a plurality of languages, the language having similar sounds and similar letters to sounds rules as a language from the plurality of languages.

20. A computer system comprising:

a processor;

a memory coupled to the processor, the memory storing instructions that when executed by the processor cause the system to perform the operations of:

generating a set of estimated pronunciations from text spellings of a set of acoustic training words, each pronunciation comprising a grouping of subword units at least some of which are common to two or more languages, the set of acoustic training words comprising at least a first word and a second word, the first word having a pronunciation based on utterances of native speakers of a first language, the second word having a pronunciation based on utterances of native speakers of a second language;

mapping sequences of sound associated with utterances of each of the acoustic training words against the estimated pronunciation associated with each of the acoustic training words;

using the mapping of sequences of sound to estimated pronunciations to generate a single acoustic subword model for each of the subword units in the grouping of subwords, the acoustic subword model comprising a sound model and a subword unit; and

using the single acoustic subword model for speech recognition.

Assignments (9)