SPEECH ENHANCEMENT METHOD, SPEECH RECOGNITION METHOD, CLUSTERING METHOD AND DEVICE

• US 20160358599A1
• Filed: 06/03/2016
• Published: 12/08/2016
• Est. Priority Date: 06/03/2015
• Status: Abandoned Application

• Current Assignee: Le Shi Zhi Xin Electronic Technology (Tianjin) Limited
• Original Assignee: Le Shi Zhi Xin Electronic Technology (Tianjin) Limited
• Inventors: WANG, Yujun
• Application Number:
  US15/173,579
• Publication Number:
  US 20160358599A1
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/20 Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech G10L21/02 takes precedence
• CPC Class Codes: G10L 2015/0633 using lexical or orthographic knowledge sources

Abstract

The present invention discloses a speech enhancement method, a speech recognition method, a clustering method and a device. The method includes: selecting a feature vector clustering center best matched with the feature vector of a first frame speech part of a test speech; performing direct to the feature vectors of other frame speech parts contained in the test speech: selecting a feature vector clustering center best matched with the feature vector of the speech part from a feature vector clustering center best matched with the feature vector of a previous frame speech part to the speech part and a feature vector clustering center adjacent to the feature vector clustering center best matched with the feature vector of the previous frame speech part; and reconstructing the feature vector of the test speech according to the feature vectors of each frame speech part contained in the test speech and the selected feature vector clustering center. Because a feature capable of representing speech continuity is utilized during speech enhancement, the present invention can achieve a better speech enhancement effect relative to a traditional speech enhancement model in the prior art.

Claims

1. A speech enhancement method, comprising:

selecting a feature vector clustering center best matched with the feature vector of a first frame speech part contained in a test speech from feature vector clustering centers obtained by training by a selection unit;

performing direct to the feature vectors of other frame speech parts contained in the test speech;

selecting a feature vector clustering center best matched with the feature vector of the speech part from a feature vector clustering center best matched with the feature vector of a previous frame speech part to the speech part and obtained by training and a feature vector clustering center adjacent to the feature vector clustering center best matched with the feature vector of the previous frame speech part, wherein a set formed by each of the feature vector clustering centers obtained by training and at least one adjacent feature vector clustering center thereof has an ability to describe speech continuity; and

reconstructing the feature vector of the test speech according to the feature vectors of each frame speech part contained in the test speech and the selected feature vector clustering center by a reconstruction unit; and

performing speech recognition on a the reconstructed feature vector of the test speech by a speech recognition.

2. The method according to claim 1, wherein reconstructing the feature vector of the test speech according to the feature vectors of each frame speech part contained in the test speech and the selected feature vector clustering center comprises:

performing an interpolation operation on a vector set formed by the feature vectors of all the speech parts contained in the test speech according to the selected feature vector clustering center, so as to obtain the reconstructed feature vector of the test speech.

3. The method according to claim 1, wherein the method, before selecting the feature vector clustering center best matched with the feature vector of the first frame speech part contained in the test speech from the feature vector clustering center obtained by training, further comprises:

respectively extracting feature vector samples from each frame speech part contained in a training corpus;

determining the distribution information of the feature vector samples in a multidimensional space;

determining initial clustering centers according to the distribution information;

performing iterative clustering on each initial clustering center to obtain undetermined clustering centers according to the similarity between the feature vector samples and each initial clustering center; and

performing iterative clustering on the undetermined clustering centers to obtain a feature vector clustering center according to given iterative clustering rules;

wherein, the given iterative clustering rules comprise;

performing iterative clustering on the undetermined clustering centers according to the feature vectors of each speech part of the training corpus;

the feature vector pursuant when performing single iterative clustering on the undetermined clustering centers being the feature vector of single speech part in the training corpus; and

the feature vectors respectively pursuant when performing every two adjacent iterative clustering on the undetermined clustering centers being the feature vectors of adjacent speech parts in the training corpus.

4. The method according to claim 3, wherein performing iterative clustering on the undetermined clustering centers to obtain the feature vector clustering center according to the given iterative clustering rules comprises:

performing iterative clustering operation direct to each training corpus according to the given iterative clustering rules, and when an iterative convergence condition is satisfied, determining each undetermined clustering center having the parameter value calculated when the iterative convergence condition is satisfied as the feature vector clustering center, wherein the iterative clustering operation comprises the following steps;

determining the similarity between the feature vector of the first frame speech part of the training corpus and the undetermined clustering center best matched with the feature vector of the first frame speech part, and the similarity between the feature vector of the first frame speech part and the undetermined clustering center adjacent to the best matched undetermined clustering center;

performing direct to other frame speech parts of the training corpus;

determining the undetermined clustering center best matched with the speech part, and determining the similarity between the feature vector of the speech part and the best matched undetermined clustering center and the similarity between the feature vector of the speech part and the undetermined clustering center adjacent to the best matched undetermined clustering center from the undetermined clustering center best matched with the feature vector of the previous frame speech part adjacent to the speech part and the clustering center adjacent to the undetermined clustering center best matched with the feature vector of the previous frame speech part adjacent to the speech part in the specific space; and

calculating the parameter values of each undetermined clustering center according to each similarity determined.

5-12. -12. (canceled)

13. An electrical apparatus, comprising:

a processor; and

an memory for storing commands executed by the processor;

wherein the processor is configured to;

selecting a feature vector clustering center best matched with the feature vector of a first frame speech part contained in a test speech from feature vector clustering centers obtained by training;

performing direct to the feature vectors of other frame speech parts contained in the test speech;

selecting a feature vector clustering center best matched with the feature vector of the speech part from a feature vector clustering center best matched with the feature vector of a previous frame speech part to the speech part and obtained by training and a feature vector clustering center adjacent to the feature vector clustering center best matched with the feature vector of the previous frame speech part, wherein a set formed by each of the feature vector clustering centers obtained by training and at least one adjacent feature vector clustering center thereof has an ability to describe speech continuity;

reconstructing the feature vector of the test speech according to the feature vectors of each frame speech part contained in the test speech and the selected feature vector clustering center; and

performing speech recognition on the reconstructed feature vector of the test speech.

15. The apparatus according to claim 13, wherein the processor is configured to:

performing an interpolation operation on a vector set formed by the feature vectors of all the speech parts contained in the test speech according to the selected feature vector clustering center, so as to obtain the reconstructed feature vector of the test speech.

16. The apparatus according to claim 13, wherein the processor is configured to:

respectively extracting feature vector samples from each frame speech part contained in a training corpus;

determining the distribution information of the feature vector samples in a multidimensional space;

determining initial clustering centers according to the distribution information;

performing iterative clustering on each initial clustering center to obtain undetermined clustering centers according to the similarity between the feature vector samples and each initial clustering center; and

performing iterative clustering on the undetermined clustering centers to obtain a feature vector clustering center according to given iterative clustering rules;

wherein, the given iterative clustering rules comprise;

performing iterative clustering on the undetermined clustering centers according to the feature vectors of each speech part of the training corpus;

the feature vector pursuant when performing single iterative clustering on the undetermined clustering centers being the feature vector of single speech part in the training corpus; and

the feature vectors respectively pursuant when performing every two adjacent iterative clustering on the undetermined clustering centers being the feature vectors of adjacent speech parts in the training corpus.

17. The apparatus according to claim 13, wherein the processor is configured to:

performing iterative clustering operation direct to each training corpus according to the given iterative clustering rules, and when an iterative convergence condition is satisfied, determining each undetermined clustering center having the parameter value calculated when the iterative convergence condition is satisfied as the feature vector clustering center, wherein the iterative clustering operation comprises the following steps;

determining the similarity between the feature vector of the first frame speech part of the training corpus and the undetermined clustering center best matched with the feature vector of the first frame speech part, and the similarity between the feature vector of the first frame speech part and the undetermined clustering center adjacent to the best matched undetermined clustering center;

performing direct to other frame speech parts of the training corpus;

determining the undetermined clustering center best matched with the speech part, and determining the similarity between the feature vector of the speech part and the best matched undetermined clustering center and the similarity between the feature vector of the speech part and the undetermined clustering center adjacent to the best matched undetermined clustering center from the undetermined clustering center best matched with the feature vector of the previous frame speech part adjacent to the speech part and the clustering center adjacent to the undetermined clustering center best matched with the feature vector of the previous frame speech part adjacent to the speech part in the specific space; and

calculating the parameter values of each undetermined clustering center according to each similarity determined.

14. (canceled)

18. A non-transitory computer storage media having computer-executable instructions stored thereon which, when executed by a computer, cause the computer to:

respectively extracting feature vector samples from each frame speech part contained in a training corpus;

determining the distribution information of the feature vector samples in a multidimensional space;

determining initial clustering centers according to the distribution information;

performing iterative clustering on each initial clustering center to obtain undetermined clustering centers according to the similarity between the feature vector samples and each initial clustering center; and

performing iterative clustering on the undetermined clustering centers to obtain a feature vector clustering center according to given iterative clustering rules;

wherein, the given iterative clustering rules comprise;

performing iterative clustering on the undetermined clustering centers according to the feature vectors of each speech part of the training corpus;

the feature vector pursuant when performing single iterative clustering on the undetermined clustering centers being the feature vector of single speech part in the training corpus; and

the feature vectors respectively pursuant when performing every two adjacent iterative clustering on the undetermined clustering centers being the feature vectors of adjacent speech parts in the training corpus.