No loss-optimization for weighted transducer

Patent metadata

Field	Value
Publication number	US-9972314-B2
Application number	US-201615170756-A
Country	US
Kind code	B2
Filing date	Jun 1, 2016
Priority date	Jun 1, 2016
Publication date	May 15, 2018
Grant date	May 15, 2018

Abstract

Techniques and architectures may be used to generate and perform a process using weighted finite-state transducers involving generic input search graphs. The process need not pursue theoretical optimality and instead search graphs may be optimized without an a priori optimization step. The process may result in an automatic speech recognition (ASR) decoder that is substantially faster than ASR decoders the include the optimization step.

First claim

What is claimed is:

1. A system comprising:
a non-transitory, computer-readable media having computer-executable instructions stored thereon; and
one or more hardware processors in communication with the non-transitory, computer-readable media that, having executed the computer-executable instructions are configured to:
instantiate a feature extractor for extracting features of input speech signals and applying sequences of one or more labels to the extracted features;
instantiate a non-optimized composed weighted speech transducer, the non-optimized composed weighted speech transducer being stored in the computer-readable media and being based on signals representing a hidden Markov model (HMM) transducer, a context dependent phone model, a lexicon model for pronunciation dictionary, and an N-gram language model for sentence probability; and
instantiate a decoder for outputting decisions about the input speech signals based, at least in part, on the sequences of labels and the non-optimized composed weighted speech transducer.

2. The system of claim 1 , wherein the non-optimized composed weighted speech transducer is configured to generate search graphs that are sub-optimal as a result of the non-optimized composed weighted speech transducer being generated by a minimization process performed on a non-deterministic transducer.

3. The system of claim 1 , wherein each of the (i) HMM transducer, (ii) the context dependent phone model, (iii) the lexicon model for pronunciation dictionary, and (iv) the N-gram language model for sentence probability respectively represent input search graphs.

4. The system of claim 1 , wherein the non-optimized composed weighted speech transducer is configured to generate a search graph.

5. The system of claim 1 , wherein the feature extractor is configured to operate online and synchronously with the input speech signals and the non-optimized composed weighted speech transducer is produced offline and asynchronously with the input speech signals.

6. A method comprising:
extracting features of input speech signals, by a feature extractor implemented by at least one hardware processor, and applying sequences of one or more labels to the extracted features;
implementing, by the at least one hardware processor, a non-optimized composed weighted speech transducer based on signals representing a hidden Markov model (HMM) transducer, a context dependent phone model, a lexicon model for pronunciation dictionary, and an N-gram language model for sentence probability; and
outputting decisions, by a decoder implemented by the at least one hardware processor, about the input speech signals based, at least in part, on the sequences of labels and the non-optimized composed weighted speech transducer.

7. The method of claim 6 , further comprising:
generating search graphs by the non-optimized composed weighted speech transducer.

8. The method of claim 7 , wherein the search graphs are sub-optimal as a result of the non-optimized composed weighted speech transducer being generated by a minimization process performed on a non-deterministic transducer.

9. The method of claim 6 , wherein each of the (i) HMM transducer, (ii) the context dependent phone model, (iii) the lexicon model for pronunciation dictionary, and (iv) the N-gram language model for sentence probability respectively represent input search graphs.

10. The method of claim 6 , wherein:
the feature extractor is configured to operate online and synchronously with the input speech signals; and
the non-optimized composed weighted speech transducer is produced offline and asynchronously with the input speech signals.

11. A non-transitory, computer-readable medium having computer-executable instructions stored thereon that, when executed by one or more hardware processors, cause the one or more hardware processors to perform a plurality of operations comprising:
extracting features of input speech signals, by an instantiated feature extractor, and applying sequences of one or more labels to the extracted features;
implementing a non-optimized composed weighted speech transducer based on signals representing a hidden Markov model (HMM) transducer, a context dependent phone model, a lexicon model for pronunciation dictionary, and an N-gram language model for sentence probability; and
outputting decisions, by an instantiated decoder, about the input speech signals based, at least in part, on the sequences of labels and the non-optimized composed weighted speech transducer.

12. The non-transitory, computer-readable medium of claim 11 , wherein the plurality of operations further comprise:
generating search graphs by the non-optimized composed weighted speech transducer.

13. The non-transitory, computer-readable medium of claim 12 , wherein the search graphs are sub-optimal as a result of the non-optimized composed weighted speech transducer being generated by a minimization process performed on a non-deterministic transducer.

14. The non-transitory, computer-readable medium 26 , wherein each of the (i) HMM transducer, (ii) the context dependent phone model, (iii) the lexicon model for pronunciation dictionary, and (iv) the N-gram language model for sentence probability respectively represent input search graphs.

15. The non-transitory, computer-readable medium of claim 11 , wherein:
the feature extractor is configured to operate online and synchronously with the input speech signals; and
the non-optimized composed weighted speech transducer is produced offline and asynchronously with the input speech signals.

Assignees

• Microsoft Technology Licensing Llc

Inventors

• Wang Zhenghao
• Huang Xuedong
• Wang Huaming

Classifications

• G06F40/284

  Lexical analysis, e.g. tokenisation or collocates · CPC title

• G10L15/193 Primary

  Formal grammars, e.g. finite state automata, context free grammars or word networks · CPC title

• G10L15/30

  Distributed recognition, e.g. in client-server systems, for mobile phones or network applications · CPC title

• G10L15/183 Primary

  using context dependencies, e.g. language models · CPC title

• G06F40/289

  Phrasal analysis, e.g. finite state techniques or chunking · CPC title