Feature normalization inputs to front end processing for automatic speech recognition

Patent metadata

Field	Value
Publication number	US-9984676-B2
Application number	US-201214414793-A
Country	US
Kind code	B2
Filing date	Jul 24, 2012
Priority date	Jul 24, 2012
Publication date	May 29, 2018
Grant date	May 29, 2018

Abstract

A computer-implemented method is described for front end speech processing for automatic speech recognition. A sequence of speech features which characterize an unknown speech input is received with a computer process. A first subset of the speech features is normalized with a computer process using a first feature normalizing function. A second subset of the speech features is normalized with a computer process using a second feature normalizing function different from the first feature normalizing function. The normalized speech features in the first and second subsets are combined with a computer process to produce a sequence of mixed normalized speech features for automatic speech recognition.

First claim

What is claimed is:

1. A method comprising:
receiving, by a computing device, a sequence of speech features that characterize an unknown speech input, the sequence of speech features including speech portions and non-speech portions;
determining, using voice activity detection (VAD), the speech portions of the sequence of speech features;
determining C 0 for the sequence of speech features, wherein C 0 comprises an average log-energy of a given speech frame of the sequence of speech features;
normalizing selected speech features of the sequence of speech features using a plurality of different feature normalizing functions including a first function used only for the speech portions and a second function used only for the non-speech portions, wherein the first function normalizes the C 0 by subtracting a max C 0 value calculated over the speech portions of the sequence of speech features, wherein the max C 0 value is initially estimated at a start of an utterance including the given speech frame and subsequently updated throughout the utterance; and
concatenating the normalized C 0 with the sequence of speech features to produce a sequence of mixed normalized speech features for automatic speech recognition of the speech portions of the sequence of speech features.

2. The method of claim 1 , further comprising:
performing automatic speech recognition of the sequence of mixed normalized speech features.

3. The method of claim 1 , wherein at least one feature normalizing function of the plurality of different feature normalizing functions is a cepstral mean normalizing function.

4. The method of claim 1 , wherein at least one feature normalizing function of the plurality of different feature normalizing functions is a cepstral variance normalizing function.

5. The method of claim 1 , wherein the selected speech features comprise a speech energy feature and at least one feature normalizing function of the plurality of different feature normalizing functions normalizes the speech energy feature.

6. The method of claim 1 , comprising:
combining the normalized selected speech features using linear discriminant analysis (LDA) to produce the sequence of mixed normalized speech features.

7. The method of claim 1 , comprising:
combining the normalized selected speech features in the feature space using neural networks to produce the sequence of mixed normalized speech features.

8. The method of claim 1 , comprising:
combining the normalized selected speech features in the feature space using feature-based minimum phone error (fMPE) to produce the sequence of mixed normalized speech features.

9. The method of claim 1 , comprising:
combining the normalized selected speech features in the model space using confusion network combination (CNC) to produce the sequence of mixed normalized speech features.

10. One or more non-transitory computer-readable media storing executable instructions that, when executed by a processor, cause a device to:
receive a sequence of speech features that characterize an unknown speech input, the sequence of speech features including speech portions and non-speech portions;
determine, using voice activity detection (VAD), the speech portions of the sequence of speech features;
determine a mean C 0 value for the sequence of speech features, wherein C 0 comprises an average log-energy of a given speech frame of the sequence of speech features,
normalize selected speech features of the sequence of speech features using a plurality of different feature normalizing functions including a first function used only for the speech portions and a second function used only for the non-speech portions, wherein the first function normalizes the C 0 by subtracting a max C 0 that is estimated at a start of an utterance including the given speech frame and subsequently updated throughout the utterance; and
combine the normalized selected speech features to produce a sequence of mixed normalized speech features for automatic speech recognition of the speech portions of the sequence of speech features.

11. The non-transitory computer-readable media of claim 10 , storing executable instructions that, when executed by the processor, cause the device to:
perform automatic speech recognition of the sequence of mixed normalized speech features.

12. The non-transitory computer-readable media of claim 10 , wherein at least one feature normalizing function of the plurality of different feature normalizing functions is a cepstral mean normalizing function.

13. The non-transitory computer-readable media of claim 10 , wherein at least one feature normalizing function of the plurality of different feature normalizing functions is a cepstral variance normalizing function.

14. The non-transitory computer-readable media of claim 10 , wherein at least one feature normalizing function of the plurality of different feature normalizing functions is a maximum normalizing function.

15. The non-transitory computer-readable media of claim 10 , wherein the selected speech features comprise a speech energy feature and at least one feature normalizing function of the plurality of different feature normalizing functions normalizes the speech energy feature.

16. The non-transitory computer-readable media of claim 10 , wherein the normalized selected speech features are combined using linear discriminant analysis (LDA) to produce the sequence of mixed normalized speech features.

17. A system comprising:
at least one processor; and
non-transitory memory storing computer-readable instructions that, when executed by the at least one processor, cause the system to:
receive a sequence of speech features that characterize an unknown speech input, the sequence of speech features including speech portions and non-speech portions;
determine, using voice activity detection (VAD), the speech portions of the sequence of speech features;
determine C 0 for the sequence of speech features, wherein C 0 comprises an average log-energy of a given speech frame of the sequence of speech features,
determine a mean C 0 value for the sequence of speech features;
normalize selected speech features of the sequence of speech features using a plurality of different feature normalizing functions including a first function used only for the speech portions and a second function used only for the non-speech portions, wherein the first function normalizes the C 0 by subtracting a max C 0 value calculated over the speech portions of the sequence of speech features, wherein the max C 0 value is initially estimated at a start of an utterance including, the given speech frame and subsequently updated throughout the utterance; and
concatenate the normalized C 0 with the sequence of speech features to produce a sequence of mixed normalized speech features for automatic speech recognition of the speech portions of the sequence of speech features.

18. The system of claim 17 , wherein the non-transitory memory stores further computer-readable instructions that, when executed by the at least one processor, cause the system to:
perform automatic speech recognition of the sequence of mixed normalized speech features.

19. The system of claim 17 , wherein at least one feature normalizing function of the plurality of different feature normalizing functions is one of a cepstral mean normalizing function and a cepstral variance normalizing function.

20. The system of claim 17 , wherein the non-transitory memory stores further computer-readable inst

Assignees

• Connolly Dermot
• Willett Daniel
• Nuance Communications Inc

Inventors

• Connolly Dermot
• Willett Daniel

Classifications

• G10L15/24

  Speech recognition using non-acoustical features · CPC title

• G10L15/02 Primary

  Feature extraction for speech recognition; Selection of recognition unit · CPC title

• G10L25/24

  the extracted parameters being the cepstrum · CPC title