Creation and use of application-generic class-based statistical language models for automatic speech recognition

• US 8,135,578 B2
• Filed: 08/24/2007
• Issued: 03/13/2012
• Est. Priority Date: 08/24/2007
• Status: Active Grant

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Hébert, Matthieu
• Primary Examiner(s): COLUCCI, MICHAEL C
• Application Number:
  US11/845,015
• Publication Number:
  US 20090055184A1
• Time in Patent Office: 1,663 Days
• Field of Search: 704/9, 704/1, 704/10, 704/231, 704/244, 704/257, 704/270.1
• US Class Current: 704/9
• CPC Class Codes: G06F 40/205 Parsing
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/193 Formal grammars, e.g. finite state automata, context free grammars or word networks
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams
• CPC Class Codes: G10L 2015/228 of application context

Abstract

A method of creating an application-generic class-based SLM includes, for each of a plurality of speech applications, parsing a corpus of utterance transcriptions to produce a first output set, in which expressions identified in the corpus are replaced with corresponding grammar tags from a grammar that is specific to the application. The method further includes, for each of the plurality of speech applications, replacing each of the grammar tags in the first output set with a class identifier of an application-generic class, to produce a second output set. The method further includes processing the resulting second output sets with a statistical language model (SLM) trainer to generate an application-generic class-based SLM.

Claims

1. A method comprising:

accessing a corpus of terms by a parser using a processor in each of a plurality of speech applications;

parsing, using said parser and said processor, said corpus of terms in each speech application to produce a plurality of first output sets, in which expressions identified in the corpus are replaced with corresponding grammar tags from a grammar that is specific to the application, wherein said grammar tags are selected from among command grammar tags and collection grammar tags;

accessing said plurality of first output sets by a class-relabeler and said processor;

replacing by the class-relabeler and said processor, for each of the plurality of speech applications, each of the grammar tags in the plurality of first output sets with a class identifier of an application-generic class, to produce plurality of a second output sets;

accessing said plurality of second output sets by a token selector and said processor;

processing collectively, by said token selector and said processor, the plurality of second output sets or data derived from the output sets with a statistical language model (SLM) trainer; and

generating, using said processor, an application-generic class-based SLM using a set of results from said SLM trainer.

2. A method as recited in claim 1, wherein the application-generic class-based

SLM includes one or more of said class identifiers.

3. A method as recited in claim 2, further comprising:

creating, using said processor, an application-specific class-based SLM for a target speech application by replacing each said class identifier in the application-generic class-based SLM with a pointer to an application-specific grammar for the target speech application.

4. A method as recited in claim 3, wherein said application-specific grammar is a class of the SLM.

5. A method as recited in claim 1, wherein said parsing comprises:

for each identified expression, identifying, using said processor, a type of grammar to which the expression corresponds; and

selecting a grammar tag to replace the expression based on the identified type of grammar.

6. A method as recited in claim 5, wherein said identifying a type of grammar comprises determining, using said processor, whether the expression corresponds to a command grammar or a collection grammar.

7. A method as recited in claim 6, wherein said replacing each of the grammar tags in the first output set with a class identifier of an application generic class comprises:

replacing, using said processor, a grammar tag with a first class identifier if the grammar tag is determined to correspond to a first type of grammar; and

replacing, using said processor, a grammar tag with a second class identifier if the grammar tag is determined to correspond to a second type of grammar.

8. A method as recited in claim 7, wherein the first type of grammar is a command grammar and the second type of grammar is a collection grammar.

9. A method as recited in claim 1, further comprising:

prior to said processing, performing on the second output sets collectively at least one operation from the set of operations consisting of;

balancing, using said processor, between the second output sets according to a size of the corpus of the corresponding speech applications;

filtering, using said processor, the second output sets to remove expressions that are not present in the corpus of at least a predetermined subset of the plurality of speech applications; and

assigning, using said processor, weights to tokens in the second output sets.

10. A method as recited in claim 1, further comprising:

executing, using said processor, an automatic speech recognition (ASR) process to recognize speech represented in a stored set of audio data associated with a target speech application, by using an application-specific grammar for the target;

speech application in combination with the application-generic class-based SLM, to generate a set of recognition results.

11. A method as recited in claim 10, further comprising:

processing, using said processor, at least a portion of the set of recognition results with an SLM trainer to generate a word-based SLM for use in ASR for the target speech application.

12. A method as recited in claim 11, further comprising:

using the word-based SLM to perform ASR for the target speech application.

13. A method of creating a statistical language model (SLM) for automatic speech recognition (ASR), the method comprising:

for each of a plurality of speech applications, parsing a corpus of utterance transcriptions, with a parser and a processor, from the application to produce a first output set, in which expressions identified in the corpus are replaced with corresponding grammar tags from a grammar that is specific to the application wherein said grammar tags are selected from among command grammar tags and collection grammar tags, wherein said parsing includes;

for each identified expression, identifying, with said parser and said processor, a type of grammar to which the expression corresponds, including determining whether the expression corresponds to a first grammar or a second grammar, and selecting, with said parser, a grammar tag to replace the expression based on the identified type of grammar;

for each of the plurality of speech applications, replacing, with a class relabeler and said processor, each of the grammar tags in the first output set with a class identifier of an application-generic class, to produce a second output set, including;

replacing, with said class-relabeler and said processor, the grammar tag with a first class identifier if the grammar tag is determined to correspond to a grammar of the first type, and

replacing, with said class-relabeler and said processor, the grammar tag with a second class identifier if the grammar tag is determined to a grammar of the second type;

filtering, with a token selector and said processor, the second output sets collectively based on an algorithm to produce a third output set; and

processing the third output set with an SLM trainer and said processor; and

generating an application-generic class-based SLM for ASR using a set of results from said SLM trainer and said processor, wherein the application-generic class-based SLM includes one or more of said class identifiers.

14. A method as recited in claim 13, wherein the first type of grammar is a command grammar and the second type of grammar is a collection grammar.

15. A method as recited in claim 14, wherein said filtering comprises at least one operation from the set of operations consisting of:

balancing, using said processor, between the second output sets according to a size of the corpus of the corresponding speech applications;

filtering, using said processor, the second output sets to remove expressions that are not present in the corpus of at least a predetermined subset of the plurality of speech applications;

assigning, using said processor, weights to tokens in the second output sets.

16. A method as recited in claim 13, further comprising:

creating, using said processor, an application-specific class-based SLM for ASR for a target application by replacing each said class identifier in the application-generic classbased SLM with a reference to an application-specific grammar for the target application.

17. A method as recited in claim 16, wherein said application-specific grammar is a class of the SLM.

18. A method as recited in claim 16, further comprising generating a word-based SLM for use in ASR, for the target speech application, by:

executing, using said processor, an ASR process to recognize speech represented in a stored set of audio data associated with the target speech application, by using said application-specific class-based SLM to generate a set of recognition results; and

processing, using said processor, at least a portion of the set of recognition results with an SLM trainer to generate the word-based SLM for use in ASR, for the target speech application.

19. A method as recited in claim 18, further comprising:

using the word-based SLM to perform ASR for the target speech application.

20. A method comprising:

accessing, using a processor, a corpus of terms by a parser in each of a plurality of speech applications;

parsing, using said parser and said processor, said corpus of terms in each speech application to produce a plurality of first output sets, in which expressions identified in the corpus are replaced with corresponding grammar tags from a grammar that is specific to the application, wherein said grammar tags are selected from among command grammar tags and collection grammar tags;

accessing, using said processor, said plurality of first output sets by a class-relabeler;

replacing, using said processor, by the class-relabeler, for each of the plurality of speech applications, each of the grammar tags in the plurality of first output sets with a class identifier of an application-generic class, to produce plurality of a second output sets;

accessing, using said processor, said plurality of second output sets by a token selector; and

processing, using said processor, collectively, by said token selector, the plurality of second output sets or data derived from the output sets with a statistical language model (SLM) trainer;

generating, using said processor, an application-generic class-based SLM using a set of results from said SLM trainer; and

creating, using said processor, an application-specific SLM for use in automatic speech recognition for a target speech application, by incorporating into the application generic class-based SLM an application-specific grammar for the target speech application.

21. A method as recited in claim 20, wherein the application-specific grammar is a class of the SLM.

22. A method as recited in claim 21, wherein creating an application specific SLM comprises replacing, using said processor, a generic class identifier in the application generic class-based SLM with a reference to an application-specific grammar for the target speech application.