Speaker identification device, speaker identification method, and recording medium

Patent metadata

Field	Value
Publication number	US-10249306-B2
Application number	US-201414760617-A
Country	US
Kind code	B2
Filing date	Jan 16, 2014
Priority date	Jan 17, 2013
Publication date	Apr 2, 2019
Grant date	Apr 2, 2019

Abstract

A speaker identification device includes: a primary speaker identification unit that computes, for each pre-stored registered speaker, a score that indicates the similarity between input speech and speech of the registered speakers; a similar speaker selection unit that selects a plurality of the registered speakers as similar speakers according to the height of the scores thereof; a learning unit that creates a classifier for each similar speaker by sorting the speech of a certain similar speaker among the similar speakers as a positive instance and the speech of the other similar speakers as negative instances; and a secondary speaker identification unit that computes, for each classifier, a score of the classifier with respect to the input speech, and outputs an identification result.

First claim

What is claimed is:

1. A speaker identification device comprising:
a primary speaker identification unit which computes, for each registered speaker stored in advance, a score that indicates similarity between input speech and speech of the registered speakers;
a similar speaker selection unit which selects a plurality of the registered speakers as similar speakers according to height of the scores;
a learning unit which creates a plurality of classifiers, each classifier corresponding to a different speaker of similar speakers,
wherein for each classifier, the classifier corresponds speech of the different speaker to which the classifier corresponds as a positive instance and speech of other speakers of the similar speakers as negative instances; and
a secondary speaker identification unit which computes, for each classifier, a score of the classifier with respect to the input speech and outputs an identification result.

2. The speaker identification device according to claim 1 , wherein
the learning unit stores in advance pairs of the similar speakers selected by the similar speaker selection unit in the past and classifiers created by the learning unit in the past as history, and creates a classifier only when there is a difference between the similar speakers in the history and the similar speakers selected by the similar speaker selection unit.

3. The speaker identification device according to claim 1 , wherein the similar speaker selection unit selects a preset number of the similar speakers.

4. The speaker identification device according to claim 1 , wherein the similar speaker selection unit selects the similar speakers based on a preset score threshold.

5. The speaker identification device according to claim 1 , wherein the classifier is a Support Vector Machine (SVM), and the score of the classifier is a distance from a feature point of the input speech to a classification plane.

6. A speaker identification method comprising:
computing, for each registered speaker stored in advance, a score that indicates similarity between input speech and speech of the registered speakers;
selecting a plurality of the registered speakers as similar speakers according to height of the scores;
creating a plurality of classifiers, each classifier corresponding to a different speaker of similar speakers;
for each classifier, corresponding, by the classifier, speech of the different speaker to which the classifier corresponds as a positive instance and speech of other speakers of the similar speakers as negative instances; and
computing, for each classifier, a score of the classifier with respect to the input speech and outputting an identification result.

7. A non-transitory computer readable medium that stores therein a speaker identification program that causes a computer to execute:
primary speaker identification processing of computing, for each registered speaker stored in advance, a score that indicates similarity between input speech and speech of the registered speakers;
similar speaker selection processing of selecting a plurality of the registered speakers as similar speakers according to height of the scores;
learning processing of creating a plurality of classifiers, each classifier corresponding to a different speaker of similar speakers;
for each classifier, corresponding, by the classifier, speech of the different speaker to which the classifier corresponds as a positive instance and speech of other speakers of the similar speakers as negative instances; and
secondary speaker identification processing of computing, for each classifier, a score of the classifier with respect to the input speech and outputs an identification result.

Assignees

• Nec Corp

Inventors

• Tani Masahiro
• Koshinaka Takafumi
• Onishi Yoshifumi
• Sawada Shigeru

Classifications

• G10L17/06

  Decision making techniques; Pattern matching strategies · CPC title

• G10L17/12 Primary

  Score normalisation · CPC title

• G10L17/04 Primary

  Training, enrolment or model building · CPC title