US 11,107,460 B2 · Adversarial speaker adaptation

Abstract

Embodiments are associated with a speaker-independent acoustic model capable of classifying senones based on input speech frames and on first parameters of the speaker-independent acoustic model, a speaker-dependent acoustic model capable of classifying senones based on input speech frames and on second parameters of the speaker-dependent acoustic model, and a discriminator capable of receiving data from the speaker-dependent acoustic model and data from the speaker-independent acoustic model and outputting a prediction of whether received data was generated by the speaker-dependent acoustic model based on third parameters. The second parameters are initialized based on the first parameters, the second parameters are trained based on input frames of a target speaker to minimize a senone classification loss associated with the second parameters, a portion of the second parameters are trained based on the input frames of the target speaker to maximize a discrimination loss associated with the discriminator, and the third parameters are trained based on the input frames of the target speaker to minimize the discrimination loss.

Claims20

Dependent claims

A system comprising:

a processing unit;
and
a memory storage device including program code that when executed by the processing unit causes to the system to:
train first parameters of a speaker-independent acoustic model to classify senones based on input speech frames;
initialize second parameters of a speaker-dependent acoustic model with the first parameters of the speaker-independent acoustic model;
initialize a discriminator to receive data from the speaker-dependent acoustic model and data from the speaker-independent acoustic model and to output a prediction of whether received data was generated by the speaker-dependent acoustic model based on third parameters of the discriminator;
determine a senone classification loss associated with the second parameters of the speaker-dependent acoustic model;
determine a discrimination loss associated with the discriminator;
train the second parameters of the speaker-dependent acoustic model based on input frames of a target speaker to minimize the senone classification loss;
train a portion of the second parameters of the speaker-dependent acoustic model based on the input frames of the target speaker to maximize the discrimination loss;
and
train the third parameters of the discriminator based on the input frames of the target speaker to minimize the discrimination loss.

9 more elements

A computer-implemented method comprising:

training first parameters of a speaker-independent acoustic model to classify senones based on input speech frames;
initializing second parameters of a speaker-dependent acoustic model with the first parameters of the speaker-independent acoustic model;
initializing a discriminator to receive data from the speaker-dependent acoustic model and data from the speaker-independent acoustic model and to output a prediction of whether received data was generated by the speaker-dependent acoustic model based on third parameters of the discriminator;
training the second parameters of the speaker-dependent acoustic model based on input frames of a target speaker to minimize a senone classification loss associated with the second parameters of the speaker-dependent acoustic model;
training a portion of the second parameters of the speaker-dependent acoustic model based on the input frames of the target speaker to maximize a discrimination loss associated with the discriminator;
and
training the third parameters of the discriminator based on the input frames of the target speaker to minimize the discrimination loss.

4 more elements

A non-transient, computer-readable medium storing program code to be executed by a processing unit to provide:

a speaker-independent acoustic model capable of classifying senones based on input speech frames and on first parameters of the speaker-independent acoustic model;
a speaker-dependent acoustic model capable of classifying senones based on input speech frames and on second parameters of the speaker-dependent acoustic model;
a discriminator capable of receiving data from the speaker-dependent acoustic model and data from the speaker-independent acoustic model and outputting a prediction of whether received data was generated by the speaker-dependent acoustic model based on third parameters of the discriminator;
and
a training platform to:
initialize the second parameters of the speaker-dependent acoustic model based on the first parameters of the speaker-independent acoustic model;
train the second parameters of the speaker-dependent acoustic model based on input frames of a target speaker to minimize a senone classification loss associated with the second parameters of the speaker-dependent acoustic model;
train a portion of the second parameters of the speaker-dependent acoustic model based on the input frames of the target speaker to maximize a discrimination loss associated with the discriminator;
and
train the third parameters of the discriminator based on the input frames of the target speaker to minimize the discrimination loss.

7 more elements

Dependencies are inferred from the claim text.

Adversarial speaker adaptation

Life of this patent

Adversarial speaker adaptation

Life of this patent

Abstract

Claims20