Dual-Band Speech Encoding

• US 20120166186A1
• Filed: 12/23/2010
• Published: 06/28/2012
• Est. Priority Date: 12/23/2010
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Acero, Alejandro, Seltzer, Michael L., Droppo, James G. III
• Granted Patent: US 8,818,797 B2
• Time in Patent Office: Days
• US Class Current: 704/205
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 19/005 Correction of errors induced by the transmission channel, if related to the coding algorithm
• CPC Class Codes: G10L 19/20 using sound class specific coding, hybrid encoders or object based coding
• CPC Class Codes: G10L 2019/0001 Codebooks
• CPC Class Codes: G10L 21/038 using band spreading techniques

Abstract

This document describes various techniques for dual-band speech encoding. In some embodiments, a first type of speech feature is received from a remote entity, an estimate of a second type of speech feature is determined based on the first type of speech feature, the estimate of the second type of speech feature is provided to a speech recognizer, speech-recognition results based on the estimate of the second type of speech feature are received from the speech recognizer, and the speech-recognition results are transmitted to the remote entity.

Claims

1. A method comprising:

determining, based on a speech waveform, a first type of speech feature and a second type of speech feature;

determining, based on the first type of speech feature, an estimate of the second type of speech feature;

determining, based on the second type of speech feature and the estimate of the second type of speech feature, an estimation error of the estimate of the second type of speech feature;

transmitting the first type of speech feature and the estimation error to a remote entity; and

receiving, from the remote entity, data associated with a speech-based service based on the first type of speech feature.

2. The method as recited in claim 1, wherein the first type of speech feature is a wideband speech feature and the second type of speech feature is a narrowband speech feature.

3. The method as described in claim 1, further comprising determining if sufficient bandwidth is available to transmit the estimation error of the second type of speech feature and, when the sufficient bandwidth is available, transmitting the estimation error of the second type of speech feature estimate.

4. The method as recited in claim 1, further comprising encoding the first type of speech feature or the estimation error of the second type of speech feature estimate using codebook-free encoding.

5. The method as recited in claim 1, wherein determining the second type of speech feature estimate uses an affine transform.

6. The method as recited in claim 1, further comprising encoding the first type of speech feature or the estimation error of the second type speech feature estimate using adaptive differential pulse-code modulation.

7. The method as recited in claim 1, further comprising encoding the first type of speech feature with dynamic mean normalization.

8. A method comprising:

receiving a first type of speech feature from a remote entity;

determining an estimate of a second type of speech feature based on the first type of speech feature;

providing the estimate of second type of speech feature to a speech recognizer trained on the second type of speech features;

receiving, from the speech recognizer, speech-recognition results based on the estimate of the second type of speech feature; and

transmitting, to the remote entity, the speech-recognition results based on the estimate of the second type of speech feature.

9. The method as recited in claim 8, wherein the first type of speech feature is encoded with adaptive differential pulse-code modulation and further comprising decoding the first type of speech feature.

10. The method as recited in claim 8, wherein determining the estimate of second type of speech feature uses an affine transform.

11. The method as recited in claim 10, wherein parameters of the affine transform are based on a parallel set of the first type of speech features and the second type of speech features.

12. The method as recited in claim 8, further comprising providing the speech-recognition results of the speech recognizer trained on the second type of speech features to a search engine, receiving search results from the search engine, and transmitting the search results to the remote entity.

13. The method as recited in claim 8, further comprising storing the first type of speech feature for training a second speech recognizer based on the first type of speech features.

14. A method comprising:

receiving, from a remote entity, a first type speech feature and an estimation error of a second type of speech feature;

determining an estimate of a second type of speech feature based on the first type of speech feature and the estimation error of the second type of speech feature;

providing the estimate of the second type of speech feature to a speech recognizer trained on the second type of speech feature;

receiving, from the speech recognizer trained on the second type of speech feature, speech-recognition results based on the estimate of the second type of speech feature estimate; and

transmitting, to the remote entity, the speech-recognition results based on the estimate of the second type of speech feature.

15. The method as recited in claim 14, wherein determining the estimate of the second type of speech feature uses an affine transform.

16. The method as recited in claim 15, wherein parameters of the affine transform are based on a parallel set of the first type of speech features and the second type of speech features.

17. The method as recited in claim 15, wherein the parameters of the affine transform are based on a minimum mean squared error model.

18. The method as recited in claim 14, wherein determining the estimate of the second type of speech feature uses a pseudo-inverse derivation model.

19. The method as recited in claim 14, further comprising providing the speech-recognition results of the speech recognizer trained on the second type of speech feature to a search engine, receiving, from the search engine, search results based on the speech-recognition results, and transmitting the search results to the remote entity.

20. The method as recited in claim 14, further comprising storing the first type speech feature for training a second speech recognizer based on the first type of speech feature.