Method and device for voiceprint recognition

• US 9,502,038 B2
• Filed: 12/12/2013
• Issued: 11/22/2016
• Est. Priority Date: 01/28/2013
• Status: Active Grant

• Current Assignee: Tencent Technology Company Limited (Tencent Holdings Limited)
• Original Assignee: Tencent Technology Company Limited (Tencent Holdings Limited)
• Inventors: Wang, Eryu, Lu, Li, Zhang, Xiang, Liu, Haibo, Li, Lou, Rao, Feng, Lu, Duling, Yue, Shuai, Chen, Bo
• Primary Examiner(s): Jackson, Jakieda
• Application Number:
  US14/105,110
• Publication Number:
  US 20140214417A1
• Time in Patent Office: 1,076 Days
• Field of Search: 704/232
• US Class Current: 1/1
• CPC Class Codes: G10L 17/02 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 Class Codes: G10L 17/04 Training, enrolment or model building
• CPC Class Codes: G10L 17/08 Use of distortion metrics or a particular distance between probe pattern and reference templates
• CPC Class Codes: G10L 17/18 Artificial neural networks; Connectionist approaches

Abstract

A method and device for voiceprint recognition, include: establishing a first-level Deep Neural Network (DNN) model based on unlabeled speech data, the unlabeled speech data containing no speaker labels and the first-level DNN model specifying a plurality of basic voiceprint features for the unlabeled speech data; obtaining a plurality of high-level voiceprint features by tuning the first-level DNN model based on labeled speech data, the labeled speech data containing speech samples with respective speaker labels, and the tuning producing a second-level DNN model specifying the plurality of high-level voiceprint features; based on the second-level DNN model, registering a respective high-level voiceprint feature sequence for a user based on a registration speech sample received from the user; and performing speaker verification for the user based on the respective high-level voiceprint feature sequence registered for the user.

Claims

1. A method, comprising:

at a device having one or more processors and memory;

establishing a first-level Deep Neural Network (DNN) model based on unlabeled speech data, the unlabeled speech data containing no speaker labels and the first-level DNN model specifying a plurality of basic voiceprint features for the unlabeled speech data;

obtaining a plurality of high-level voiceprint features by tuning the first-level DNN model based on labeled speech data, the labeled speech data containing speech samples with respective speaker labels, and the tuning producing a second-level DNN model specifying the plurality of high-level voiceprint features;

based on the second-level DNN model, registering a first high-level voiceprint feature sequence for a user based on a registration speech sample received from the user; and

performing speaker verification for the user based on the first high-level voiceprint feature sequence registered for the user, the speaker verification comprising;

receiving, from the user, a test speech sample;

obtaining a second high-level voiceprint feature sequence based on the test speech sample using the first-level DNN model and the second-level DNN model in sequence;

determining a distance between the second high-level voiceprint feature sequence and the first high-level voiceprint feature sequence registered for the user; and

in accordance with a determination that the distance between the second high-level voiceprint feature sequence and the first high-level voiceprint feature sequence is less than a preset threshold, automatically, without user intervention, verifying the identity of the user.

2. The method of claim 1, wherein tuning the first-level DNN model based on the labeled speech data to produce the second-level DNN model comprises:

imposing at least two constraints during adjustment of the first-level DNN based on the labeled speech data, including;

(1) distances between characteristic voiceprints generated from speech samples of different speakers increase with training, and (2) distances between characteristic voice prints generated from speech samples of same speakers decrease with training.

3. The method of claim 1, wherein tuning the first-level DNN model based on the labeled speech data comprises:

applying sparse coding limit rules to train the first-level DNN model based on the labeled speech data.

4. The method of claim 1, wherein tuning the first-level DNN model based on the labeled speech data comprises:

applying maximum cross entropy rules to train the first-level DNN model based on the labeled speech data.

5. The method of claim 1, wherein, based on the second-level DNN model, registering the first high-level voiceprint feature sequence for the user based on the registration speech sample received from the user comprises:

obtaining a respective basic voiceprint feature sequence for the user from the registration speech sample based on the first-level DNN model; and

providing the respective basic voiceprint feature sequence as input to the second-level DNN model to obtain the first high-level voiceprint feature sequence for the user.

6. The method of claim 1, wherein performing speaker verification for the user based on the first high-level voiceprint feature sequence registered for the user comprises:

verifying the user'"'"'s identity based on a Euclidean distance between respective Gaussian models of the second high-level voiceprint feature sequence obtained from the test speech sample and the first high-level voiceprint feature sequence registered for the user.

7. A voiceprint recognition system, comprising:

one or more processors; and

memory storing instructions that, when executed by the one or more processors, cause the processors to perform operations comprising;

establishing a first-level Deep Neural Network (DNN) model based on unlabeled speech data, the unlabeled speech data containing no speaker labels and the first-level DNN model specifying a plurality of basic voiceprint features for the unlabeled speech data;

obtaining a plurality of high-level voiceprint features by tuning the first-level DNN model based on labeled speech data, the labeled speech data containing speech samples with respective speaker labels, and the tuning producing a second-level DNN model specifying the plurality of high-level voiceprint features;

based on the second-level DNN model, registering a first high-level voiceprint feature sequence for a user based on a registration speech sample received from the user; and

performing speaker verification for the user based on the first high-level voiceprint feature sequence registered for the user, the speaker verification comprising;

receiving, from the user, a test speech sample;

obtaining a second high-level voiceprint feature sequence based on the test speech sample using the first-level DNN model and the second-level DNN model in sequence;

determining a distance between the second high-level voiceprint feature sequence and the first high-level voiceprint feature sequence registered for the user; and

in accordance with a determination that the distance between the second high-level voiceprint feature sequence and the first high-level voiceprint feature sequence is less than a preset threshold, automatically, without user intervention, verifying the identity of the user.

8. The system of claim 7, wherein tuning the first-level DNN model based on the labeled speech data to produce the second-level DNN model comprises:

imposing at least two constraints during adjustment of the first-level DNN based on the labeled speech data, including;

(1) distances between characteristic voiceprints generated from speech samples of different speakers increase with training, and (2) distances between characteristic voice prints generated from speech samples of same speakers decrease with training.

9. The system of claim 7, wherein tuning the first-level DNN model based on the labeled speech data comprises:

applying sparse coding limit rules to train the first-level DNN model based on the labeled speech data.

10. The system of claim 7, wherein tuning the first-level DNN model based on the labeled speech data comprises:

applying maximum cross entropy rules to train the first-level DNN model based on the labeled speech data.

11. The system of claim 7, wherein, based on the second-level DNN model, registering the first high-level voiceprint feature sequence for the user based on the registration speech sample received from the user comprises:

obtaining a respective basic voiceprint feature sequence for the user from the registration speech sample based on the first-level DNN model; and

providing the respective basic voiceprint feature sequence as input to the second-level DNN model to obtain the first high-level voiceprint feature sequence for the user.

12. The system of claim 7, wherein performing speaker verification for the user based on the first high-level voiceprint feature sequence registered for the user comprises:

verifying the user'"'"'s identity based on a Euclidean distance between respective Gaussian models of the second high-level voiceprint feature sequence obtained from the test speech sample and the first high-level voiceprint feature sequence registered for the user.

13. A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the processors to perform operations comprising:

establishing a first-level Deep Neural Network (DNN) model based on unlabeled speech data, the unlabeled speech data containing no speaker labels and the first-level DNN model specifying a plurality of basic voiceprint features for the unlabeled speech data;

obtaining a plurality of high-level voiceprint features by tuning the first-level DNN model based on labeled speech data, the labeled speech data containing speech samples with respective speaker labels, and the tuning producing a second-level DNN model specifying the plurality of high-level voiceprint features;

based on the second-level DNN model, registering a first high-level voiceprint feature sequence for a user based on a registration speech sample received from the user; and

performing speaker verification for the user based on the first high-level voiceprint feature sequence registered for the user, the speaker verification comprising;

receiving, from the user, a test speech sample;

obtaining a second high-level voiceprint feature sequence based on the test speech sample using the first-level DNN model and the second-level DNN model in sequence;

determining a distance between the second high-level voiceprint feature sequence and the first high-level voiceprint feature sequence registered for the user; and

in accordance with a determination that the distance between the second high-level voiceprint feature sequence and the first high-level voiceprint feature sequence is less than a preset threshold, automatically, without user intervention, verifying the identity of the user.

14. The computer-readable medium of claim 13, wherein tuning the first-level DNN model based on the labeled speech data to produce the second-level DNN model comprises:

imposing at least two constraints during adjustment of the first-level DNN based on the labeled speech data, including;

(1) distances between characteristic voiceprints generated from speech samples of different speakers increase with training, and (2) distances between characteristic voice prints generated from speech samples of same speakers decrease with training.

15. The computer-readable medium of claim 13, wherein tuning the first-level DNN model based on the labeled speech data comprises:

applying sparse coding limit rules to train the first-level DNN model based on the labeled speech data.

16. The computer-readable medium of claim 13, wherein tuning the first-level DNN model based on the labeled speech data comprises:

applying maximum cross entropy rules to train the first-level DNN model based on the labeled speech data.

17. The computer-readable medium of claim 13, wherein, based on the second-level DNN model, registering the first high-level voiceprint feature sequence for the user based on the registration speech sample received from the user comprises:

obtaining a respective basic voiceprint feature sequence for the user from the registration speech sample based on the first-level DNN model; and

providing the respective basic voiceprint feature sequence as input to the second-level DNN model to obtain the first high-level voiceprint feature sequence for the user.