Dual model speaker identification

Patent metadata

Field	Value
Publication number	US-9711148-B1
Application number	US-201313944975-A
Country	US
Kind code	B1
Filing date	Jul 18, 2013
Priority date	Jul 18, 2013
Publication date	Jul 18, 2017
Grant date	Jul 18, 2017

Abstract

A processing system receives an audio signal encoding an utterance and determines that a first portion of the audio signal corresponds to a predefined phrase. The processing system accesses one or more text-dependent models associated with the predefined phrase and determines a first confidence based on the one or more text-dependent models associated with the predefined phrase, the first confidence corresponding to a first likelihood that a particular speaker spoke the utterance. The processing system determines a second confidence for a second portion of the audio signal using one or more text-independent models, the second confidence corresponding to a second likelihood that the particular speaker spoke the utterance. The processing system then determines that the particular speaker spoke the utterance based at least in part on the first confidence and the second confidence.

First claim

What is claimed is:

1. A computer-implemented method comprising:
receiving, by a speech-enabled home device of an automated speaker identification system that includes the speech-enabled home device that includes one or more microphones for detecting utterances spoken in a home environment, and a server based speaker recognition engine that is associated with an automated query processor and that includes (i) one or more text-dependent speaker identification models that are trained using multiple previous utterances of a keyword by a particular speaker and by other users whose accounts are registered with the server, and (ii) one or more text-independent speaker identification models that are trained using utterances of words other than the keyword by the particular speaker and by other users whose accounts are registered with the server, an audio signal encoding an utterance that was spoken in a home environment and was detected by one or more microphones of the speech-enabled home device, and that includes the keyword and a query;
determining, by the server-based speaker recognition engine and based on an analysis of a portion of the audio signal that corresponds to the keyword by one or more of the text-dependent speaker identification models that are trained using utterances of the keyword by the particular speaker and by the other users whose accounts are registered with the server, a first speaker identification confidence value that reflects a likelihood that the particular speaker spoke the keyword;
determining, by the server-based speaker recognition engine and based on an analysis of at least a portion of the audio signal that corresponds to the query by one or more of the text-independent speaker identification models that are trained using utterances of words other than the keyword by the particular speaker and by other users whose accounts are registered with the server, a second speaker identification confidence value that reflects a likelihood that the particular speaker spoke the query;
determining, by the server-based speaker recognition engine, a first quantity of the utterances of the keyword by the particular speaker that were used to train the one or more text-dependent speaker identification models;
determining, by the server-based speaker recognition engine, a second quantity of the utterances of the words other than the keyword by the particular speaker that were used to train the one or more text-independent speaker identification models;
assigning, by the server-based speaker recognition engine, a first weight to the first speaker identification confidence value based at least on the first quantity of utterances of the keyword by the particular speaker that were used to train the one or more text-dependent speaker identification models, and a second weight to the second speaker identification confidence value based at least on the second quantity of utterances of the words other than the keyword by the particular speaker that were used to train the one or more text-independent speaker identification models;
determining, by the server-based speaker recognition engine, that the particular speaker spoke the utterance encoded in the audio signal based at least in part on the weighted first speaker identification confidence value and the weighted second speaker identification confidence value;
in response to determining by the server-based speaker recognition engine that the particular speaker spoke the utterance encoded in the audio signal based at least in part on the weighted first speaker identification confidence value and the weighted second speaker identification confidence value, initiating access to one or more account resources associated with the particular speaker for preparation of a personalized response to the query by the automated query processor that is associated with the server-based speaker recognition engine; and
providing, by the automated query processor that is associated with the server-based speaker recognition engine, the personalized response to the speech-enabled home device for output to the particular speaker.

2. The method of claim 1 , comprising:
obtaining one or more sets of mel-frequency cepstral coefficients (MFCCs) associated with the keyword, each set of MFCCs being associated with an individual speaker; and
wherein determining, based on an analysis of a portion of the audio signal that corresponds to the keyword by one or more of the more text-dependent speaker identification models that are trained using utterances of the keyword by the particular speaker, the first speaker identification confidence value that reflects the likelihood that the particular speaker spoke the keyword, comprises determining, based on a comparison of the one or more sets of MFCCs to a set of MFCCs derived from the portion of the audio signal that corresponds to the keyword, a first speaker identification confidence value that reflects a likelihood that the particular speaker spoke the keyword.

3. The method of claim 1 , wherein determining the second speaker identification confidence value comprises:
deriving a set of mel-frequency cepstral coefficients (MFCCs) from the portion of the audio signal that corresponds to the query;
accessing one or more Gaussian mixture models (GMMs), each GMM being associated with an individual speaker; and
processing the set of MFCCs from the portion of the audio signal that corresponds to the query using each of the GMMs to determine the second speaker identification confidence value.

4. The method of claim 1 further comprising:
analyzing the portion of the audio signal that corresponds to the keyword using the one or more text-independent models to determine a third speaker identification confidence value that reflects a likelihood that the particular speaker generated the utterance; and
wherein determining that the particular speaker spoke the utterance based at least in part on the weighted first confidence and the weighted second confidence comprises determining that the particular speaker spoke the utterance based at least in part on the weighted first confidence, the weighted second confidence, and the third speaker identification confidence value.

5. The method of claim 1 , wherein determining that the particular speaker spoke the utterance based at least on the weighted first speaker identification confidence value and the weighted second speaker identification confidence value comprises:
combining the weighted first speaker identification confidence value and the weighted second speaker identification confidence value to generate a combined confidence; and
determining that the combined confidence for the particular speaker is greater than a combined confidence for any other speaker.

6. The method of claim 5 , wherein determining that the combined confidence for the particular speaker is greater than a combined confidence for any other speaker comprises determining that the combined confidence for the particular speaker is greater than a combined confidence for any other speaker and that the combined confidence satisfies a predetermined threshold.

7. The method of claim 1 , wherein determining that the particular speaker spoke the utterance based at least in part on the weighted first speaker identification confidence value and the weighted second speaker identification confidence value comprises determining that the particular speaker from among a plurality of speakers spoke the utterance based at least in part on the weighted first speaker identification confidence value and the weighted second speaker identification confidence value.

8. The method of claim 1 , further comprising:
combining the weighted first speaker identification confidence value and the weighted sec

Assignees

• Google Inc

Inventors

• Sharifi Matthew
• Roblek Dominik

Classifications

• G10L17/10 Primary

  Multimodal systems, i.e. based on the integration of multiple recognition engines or fusion of expert systems · CPC title

• G10L17/02 Primary

  Preprocessing operations, e.g. segment selection; Pattern representation or modelling, e.g. based on linear discriminant analysis [LDA] or principal components; Feature selection or extraction · CPC title

• G10L17/22

  Interactive procedures; Man-machine interfaces · CPC title

• G10L15/02

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

• G10L17/04

  Training, enrolment or model building · CPC title