# US 2013/0253931 A1 · Modeling device and method for speaker recognition, and speaker recognition system

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2013/0253931 A1
- Document ID: US20130253931A1
- Status: Published application
- Kind: A1
- Application number: 13/989,508
- Filed: 2010-12-10
- Published: 2013-09-26
- Earliest claimed: 2010-12-10
- Assignees: Panasonic Intellectual Property Corporation of America
- Inventors: Haifeng Shen; Long Ma; Bingqi Zhang
- Web page: https://patentgrep.com/patent/US20130253931A1

## CPC

- G10L 17/04: Training, enrolment or model building
- G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING

## Abstract

A modeling device and method for speaker recognition and a speaker recognition system are provided. The modeling device comprises a front end which receives enrollment speech data from each target speaker, a reference anchor set generation unit which generates a reference anchor set using the enrollment speech data based on an anchor space, and a voice print generation unit which generates voice prints based on the reference anchor set and the enrollment speech data. With the present disclosure, by taking the enrollment speech and speaker adaptation technique into account, anchor models with smaller size can be generated, so reliable and robust speaker recognition with smaller size reference anchor set is possible. It brings great advantages for computation speed improvement and great memory reduction.

## Claims

### Claim 1

A modeling device for speaker recognition, comprising:
a front end which receives enrollment speeches from target speakers;
a reference anchor set generation unit which generates a reference anchor set using the enrollment speeches based on an anchor space; and
a voice print generation unit which generates voice prints based on the reference anchor set and the enrollment speeches.

### Claim 2 (depends on 1)

The modeling device for speaker recognition according to claim 1, wherein the reference anchor set is composed of principal anchors and associate anchors, and the associate anchors are generated based on the principal anchors.

### Claim 3 (depends on 2)

The modeling device for speaker recognition according to claim 2, wherein the anchor space includes a plurality of anchor models, the principal anchors are generated by finding anchor models with the nearest distance to the models of enrollment speeches from the anchor space.

### Claim 4 (depends on 2)

The modeling device for speaker recognition according to claim 2, wherein the associate anchors are generated by splitting the anchor space into a plurality of clusters based on the principal anchors and finding anchor models with the nearest distance to the centroids of the plurality of clusters containing no principal anchors.

### Claim 5 (depends on 4)

The modeling device for speaker recognition according to claim 4, wherein the anchor space is split into the plurality of clusters by: firstly splitting the anchor space into N clusters wherein N equals to the number of the principal anchors; finding one cluster with the biggest intra-class distance among the N clusters to be further split, wherein if the cluster has a principal anchor, the cluster is split into two sub-clusters based on the principal anchor and an anchor with the farthest distance to the principal anchor in the cluster, while if the cluster has no principal anchor, the cluster is split into two sub-clusters based on two farthest anchors in the cluster with respect to the father principal anchor; and repeating the above process until the plurality of clusters are obtained.

### Claim 6 (depends on 3)

The modeling device for speaker recognition according to claim 3, wherein the distance is a likelihood value.

### Claim 7 (depends on 2)

The modeling device for speaker recognition according to claim 2, wherein the principal anchors are adapted with the enrollment speeches by adaptation method such as MAP or MLLR, and the associate anchors are generated based on the principal anchors after adaptation.

### Claim 8 (depends on 2)

The modeling device for speaker recognition according to claim 2, wherein the principal anchors are adapted with the enrollment speeches by adaptation method such as MAP or MLLR, and the associate anchors and the adapted principal anchors are combined as the reference anchor set.

### Claim 9

A modeling method for speaker recognition, comprising steps of:
receiving enrollment speeches from target speakers;
generating a reference anchor set using the enrollment speeches based on an anchor space; and
generating voice prints based on the reference anchor set and enrollment speeches.

### Claim 10 (depends on 9)

The modeling method for speaker recognition according to claim 9, wherein the reference anchor set is composed of principal anchors and associate anchors, and the method further comprising a step of generating the associate anchors based on the principal anchors.

### Claim 11 (depends on 10)

The modeling method for speaker recognition according to claim 10, wherein the anchor space includes a plurality of anchor models, and the method further comprising a step of generating the principal anchors by finding anchor models with the nearest distance to the models of enrollment speeches from the anchor space.

### Claim 12 (depends on 10)

The modeling method for speaker recognition according to claim 10, further comprising the steps of generating the associate anchors by splitting the anchor space into a plurality of clusters based on the principal anchors and finding anchor models with the nearest distance to the centroids of the plurality of clusters containing no principal anchors.

### Claim 13 (depends on 12)

The modeling method for speaker recognition according to claim 12, further comprising the steps of:
firstly splitting the anchor space into N clusters, wherein N equals to the number of the principal anchors;
finding one cluster with the biggest intra-class distance among the N clusters to be further split, wherein if the cluster has a principal anchor, the cluster is split into two sub-clusters based on the principal anchor and an anchor with the farthest distance to the principal anchor in the cluster, while if the cluster has no principal anchor, the cluster is split into two sub-clusters based on two farthest anchors in the cluster with respect to the father principal anchor; and
repeating the above process until the plurality of clusters are obtained.

### Claim 14 (depends on 11)

The modeling method for speaker recognition according to claim 11, wherein the distance is a likelihood value.

### Claim 15 (depends on 10)

The modeling method for speaker recognition according to claim 10, further comprising the steps of:
adapting the principal anchors with the enrollment speeches by adaptation method such as MAP or MLLR; and
generating the associate anchors based on the principal anchors after adaptation.

### Claim 16 (depends on 10)

The modeling method for speaker recognition according to claim 10, further comprising steps of
adapting the principal anchors with the enrollment speeches by adaptation method such as MAP or MLLR; and
combining the associate anchors and the adapted principal anchors as the reference anchor set.

### Claim 17

A speaker recognition system comprises:
a front end which receives enrollment speeches and/or testing speeches from target speakers;
a reference anchor set generation unit which generates a reference anchor set using the enrollment speeches based on an anchor space;
a voice print generation unit which generates voice prints based on the reference anchor set and the enrollment speeches and/or testing speeches;
a matching unit which compares the voice prints generated from the testing speeches with the voice prints generated from the enrollment speeches; and
a decision unit which recognizes the identity of the target speaker based on the result of the comparison.

### Claim 18 (depends on 17)

The speaker recognition system according to claim 17, wherein the reference anchor set is composed of principal anchors and associate anchors, and the associate anchors are generated based on the principal anchors.

### Claim 19 (depends on 18)

The speaker recognition system according to claim 18, wherein the anchor space includes a plurality of anchor models, the principal anchors are generated by finding anchor models with the nearest distance to the models of enrollment speeches from the anchor space.

### Claim 20 (depends on 18)

The speaker recognition system according to claim 18, wherein the associate anchors are generated by splitting the anchor space into a plurality of clusters based on the principal anchors and finding anchor models with the nearest distance to the centroids of the plurality of clusters containing no principal anchors.

### Claim 21 (depends on 20)

The speaker recognition system according to claim 20, wherein the anchor space is split into the plurality of clusters by: firstly splitting the anchor space into N clusters wherein N equals to the number of the principal anchors; finding one cluster with the biggest intra-class distance among the N clusters to be further split, wherein if the cluster has a principal anchor, the cluster is split into two sub-clusters based on the principal anchor and an anchor with the farthest distance to the principal anchor in the cluster, while if the cluster has no principal anchor, the cluster is split into two sub-clusters based on two farthest anchors in the cluster with respect to the father principal anchor; and repeating the above process until the plurality of clusters are obtained.

### Claim 22 (depends on 19)

The speaker recognition system according to claim 19, wherein the distance is a likelihood value.

### Claim 23 (depends on 18)

The speaker recognition system according to claim 18, wherein the principal anchors are adapted with the enrollment speeches by adaptation method such as MAP or MLLR, and the associate anchors are generated based on the principal anchors after adaptation.

### Claim 24 (depends on 18)

The speaker recognition system according to claim 18, wherein the principal anchors are adapted with the enrollment speeches by adaptation method such as MAP or MLLR, and the associate anchors and the adapted principal anchors are combined as the reference anchor set.

### Claim 25

A speaker recognition system comprises a modeling device and a recognition device, wherein,
the modeling device comprising:
a first front end which receives enrollment speeches from target speakers;
a reference anchor set generation unit which generates a reference anchor set using the enrollment speeches based on an anchor space; and
a first voice print generation unit which generates first voice prints based on the reference anchor set and the enrollment speeches, and
the recognition device comprising:
a second front end which receives testing speeches from the target speakers;
a second voice print generation unit which generates second voice prints based on the reference anchor set and the testing speeches;
a matching unit which compares the first voice prints with the second voice prints; and
a decision unit which recognizes the identity of the target speakers based on the result of comparison.
