Speech recognition using speech characteristic probabilities

• US 9,202,470 B2
• Filed: 01/31/2013
• Issued: 12/01/2015
• Est. Priority Date: 09/28/2009
• Status: Active Grant

• Current Assignee: Avago Technologies International Sales Pte Limited (Broadcom, Inc.)
• Original Assignee: Broadcom Corporation (Broadcom, Inc.)
• Inventors: Seshadri, Nambirajan
• Primary Examiner(s): Lerner, Martin
• Application Number:
  US13/755,975
• Publication Number:
  US 20130151254A1
• Time in Patent Office: 1,034 Days
• Field of Search: 704/8, 704/240, 704/246, 704/247, 704/249, 704/250, 704/257
• US Class Current: 1/1
• CPC Class Codes: G10L 15/005 Language recognition
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/28 Constructional details of speech recognition systems
• CPC Class Codes: G10L 25/00 Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 muting semiconductor-based amplifiers when some special characteristics of a signal are sensed by a speech detector, e.g. sensing when no signal is present, H03G3/34

Abstract

A speech recognition module includes an acoustic front-end module, a sound detection module, and a word detection module. The acoustic front-end module generates a plurality of representations of frames from a digital audio signal and generates speech characteristic probabilities for the plurality of frames. The sound detection module determines a plurality of estimated utterances from the plurality of representations and the speech characteristic probabilities. The word detection module determines one or more words based on the plurality of estimated utterances and the speech characteristic probabilities.

Claims

1. An acoustic front-end device comprising:

a frame parser receiving an acoustic signal and parsing the received acoustic signal into a plurality of frames;

a plurality of correlators, each of the correlators correlating each of the plurality of frames with one or more acoustic property sets to produce a first set of acoustic property correlations;

a controller

retrieving one or more speech characteristic samples from one or more speech codebooks based on the first set of acoustic property correlations; and

a speech characteristic probability generator configured to;

generate a plurality of speech characteristic probabilities over one or more subsequent frames of the plurality of frames by individually correlating a digital signal component of the subsequent frame with the one or more acoustic property sets to produce a second set of acoustic property correlations; and

a processor configured to interpret the plurality of speech characteristic probabilities to generate at least a language probability and a language syntax bias; and

select a plurality of words from a series of plurality of words based on the language probability and the language syntax bias; and

output the plurality of words through an interface.

2. The device of claim 1, wherein the controller retrieving one or more speech characteristic samples from the one or more speech codebooks is further configured to:

interpret a set of the acoustical property set correlations to produce a first index;

and address the one or more speech codebooks based on the first index to retrieve the one or more speech characteristic samples.

3. The device of claim 1, wherein the one or more speech codebooks store a plurality of the speech characteristic samples comprising a plurality of speech characteristics.

4. The device of claim 3, wherein the plurality of speech characteristics comprises any of:

gender, age, nationality, dialect or accent.

5. The device of claim 1, wherein at least one of the one or more speech codebooks comprises any of:

sound tables, word tables or syntax tables.

6. The device of claim 1, wherein at least one of the one or more speech codebooks is based on a hidden Markov model.

7. The device of claim 6, wherein the one or more codebooks based on the hidden Markov model provide, in sequence, a plurality of features vectors representing spectral characteristics of speech for a given frame for the plurality of frames.

8. The device of claim 1, wherein the controller analyzes each of the acoustic property set correlations to determine one or more likely speech characteristics of the frame.

9. The device of claim 8, wherein, if the controller when analyzing each of the acoustic property set correlations, for a specified speech characteristic of the one or more likely speech characteristics, cannot determine conclusively that the specified speech characteristic is present, it ignores the specified speech characteristic.

10. The device of claim 2 further comprising:

interpreting the second set of acoustical property correlations with respect to the speech characteristic probability for the subsequent frame to produce a second index;

addressing the one or more speech codebooks based on the second index to retrieve a second speech characteristic sample; and

correlating the digital signal component with the second speech characteristic sample to produce the speech characteristic probability for the subsequent frame.

11. The device of claim 10, wherein the processor is further configured to:

interpret the plurality of speech characteristic probabilities to further generate a word bias and

select a plurality of words from the series of plurality of words based on the word bias.

12. A method for frame-by-frame analyzing of an acoustic signal to determine speech characteristics comprising:

configuring a processor to;

parse the acoustic signal into a plurality of frames;

correlate each of the plurality of frames with one or more acoustic property sets to produce a first set of acoustic property correlations and first index;

retrieve one or more first speech characteristic samples from one or more speech codebooks based on the first index; and

generate a plurality of speech characteristic probabilities for a subsequent frame of the plurality of frames by;

individually correlating a digital signal component of the subsequent frame with the one or more acoustical property sets to produce a second set of acoustic property correlations

and a second index;

addressing the one or more speech codebooks based on the second index to retrieve a second speech characteristic sample; and

correlating the digital signal component with the second speech characteristic sample to produce the speech characteristic probability for the subsequent frame; and

retrieving and outputting words based on a language probability and a language syntax bias of the plurality of speech characteristic probabilities.

13. The method claim 12, wherein the retrieving one or more speech characteristic samples from the one or more speech codebooks further comprises:

interpreting a third set of acoustic property correlations to produce a third index; and

addressing the one or more speech codebooks based on the third index to retrieve one or more third speech characteristic samples.

14. The method of claim 12 further configuring the processor to analyze each of the acoustic property set correlations to determine one or more likely speech characteristics of the frame.

15. The method of claim 12, wherein the first and second speech characteristic samples comprise a plurality of speech characteristics comprising any of:

gender, age, nationality, dialect or accent.

16. The method of claim 12, wherein at least one of the one or more speech codebooks comprises any of:

sound tables, word tables or syntax tables.

17. The method of claim 12 further comprising configuring the processor to:

interpret the plurality of speech characteristic probabilities to generate a word bias;

select the plurality of words from a series of plurality of words based on the word bias; and

output the one or more words through an interface.

18. An acoustic front-end device comprising:

a frame parser receiving an acoustic signal and parsing the received acoustic signal into a plurality of frames;

a plurality of correlators, each of the correlators correlating each of the plurality of frames with the one or more acoustic property sets to produce a first set of acoustic property correlations;

a controller, the controller retrieving one or more first speech characteristic samples from one or more speech codebooks based on the acoustic property set correlations;

a speech characteristic probability generator generating speech characteristic probabilities for a subsequent frame of the plurality of frames by;

individually correlating a digital signal component of the subsequent frame with the one or more acoustical property sets to produce a second set of acoustic property correlations

addressing the one or more speech codebooks based on the second set of acoustic property correlations to retrieve a second speech characteristic sample; and

correlating the digital signal component with the second speech characteristic sample to produce the speech characteristic probability for the subsequent frame; and

a processor, the processor;

interpreting the plurality of speech characteristic probabilities to generate at least a language probability, word bias, and language syntax bias;

selecting a plurality of words from a series of plurality of words based on the word bias and language probability; and

selecting a plurality of language syntaxes from a series of plurality of language syntaxes based on the language probability.

19. The device of claim 18, wherein at least one of the one or more speech codebooks is based on a hidden Markov model.