Speaker adaptation of neural network acoustic models using I-vectors

Patent metadata

Field	Value
Publication number	US-9858919-B2
Application number	US-201414500042-A
Country	US
Kind code	B2
Filing date	Sep 29, 2014
Priority date	Nov 27, 2013
Publication date	Jan 2, 2018
Grant date	Jan 2, 2018

Abstract

A method includes providing a deep neural network acoustic model, receiving audio data including one or more utterances of a speaker, extracting a plurality of speech recognition features from the one or more utterances of the speaker, creating a speaker identity vector for the speaker based on the extracted speech recognition features, and adapting the deep neural network acoustic model for automatic speech recognition using the extracted speech recognition features and the speaker identity vector.

First claim

What is claimed is:

1. A method comprising:
providing a deep neural network acoustic model;
receiving audio data including one or more utterances of a speaker;
extracting a plurality of speech recognition features from the one or more utterances of the speaker;
creating a speaker identity vector for the speaker based on the speech recognition features extracted from the one or more utterances of the speaker;
performing, by a computer system, an automatic speech recognition using the speech recognition features extracted from the one or more utterances of the speaker and the speaker identity vector by executing the deep neural network acoustic model; and
adapting the deep neural network acoustic model executing on the computer system performing the automatic speech recognition using the speech recognition features extracted from the one or more utterances of the speaker and the speaker identity vector, wherein adapting the deep neural network acoustic model further comprises concatenating the speaker identity vector to each of the speech recognition features extracted from the one or more utterances of the speakers to form an input to the deep neural network acoustic model.

2. The method of claim 1 , wherein the speaker identity vector encapsulates information about an identity of the speaker in a low-dimensional fixed-length representation.

3. The method of claim 1 , wherein adapting the deep neural network acoustic model further comprises:
training a speaker-independent Gaussian Mixture Model; and
aligning the audio data to the speaker-independent Gaussian Mixture Model to determine zero-order statistics and first-order statistics.

4. The method of claim 1 , further comprising clustering a plurality of speakers using respective speaker identity vectors.

5. The method of claim 1 , further comprising clustering a plurality of utterances using respective speaker identity vectors.

6. A computer program product for adapting deep neural network acoustic models for automatic speech recognition, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to perform a method comprising:
providing a deep neural network acoustic model;
receiving audio data including one or more utterances of a speaker;
extracting a plurality of speech recognition features from the one or more utterances of the speaker;
creating a speaker identity vector for the speaker based on the speech recognition features extracted from the one or more utterances of the speaker;
performing, by the processor, an automatic speech recognition using the speech recognition features extracted from the one or more utterances of the speaker and the speaker identity vector by executing the deep neural network acoustic model; and
adapting the deep neural network acoustic model executing on a computer system the processor performing the automatic speech recognition using the speech recognition features extracted from the one or more utterances of the speaker and the speaker identity vector, wherein adapting the deep neural network acoustic model further comprises concatenating the speaker identity vector to each of the speech recognition features extracted from the one or more utterances of the speakers to form an input to the deep neural network acoustic model.

7. The computer program product of claim 6 , wherein the speaker identity vector encapsulates information about an identity of the speaker in a low-dimensional fixed-length representation.

8. The computer program product of claim 6 , wherein adapting the deep neural network acoustic model further comprises:
training a speaker-independent Gaussian Mixture Model; and
aligning the audio data to the speaker-independent Gaussian Mixture Model to determine zero-order statistics and first-order statistics.

9. The computer program product of claim 6 , further comprising clustering a plurality of speakers using respective speaker identity vectors.

10. The computer program product of claim 6 , further comprising clustering a plurality of utterances using respective speaker identity vectors.

Assignees

• IBM

Inventors

• Saon George A

Classifications

• G10L15/16

  using artificial neural networks · CPC title

• G10L17/18

  Artificial neural networks; Connectionist approaches · CPC title

• G10L15/063 Primary

  Training · CPC title