APPARATUS AND METHOD FOR SPEECH RECOGNITION, AND APPARATUS AND METHOD FOR TRAINING TRANSFORMATION PARAMETER

• US 20170084268A1
• Filed: 09/13/2016
• Published: 03/23/2017
• Est. Priority Date: 09/18/2015
• Status: Abandoned Application

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: Samsung Electronics Co. Ltd.
• Inventors: Yoo, Sang Hyun, Kim, Jung Hoe, PARK, Hyoung Min, Choi, Hee Youl, Hong, Seok Jin
• Application Number:
  US15/263,977
• Publication Number:
  US 20170084268A1
• Time in Patent Office: Days
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/07 to the speaker
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 15/14 using statistical models, e.g. Hidden Markov Models [HMMs] G10L15/18 takes precedence
• CPC Class Codes: G10L 15/18 using natural language modelling
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 2015/0635 updating or merging of old and new templates; Mean values; Weighting
• CPC Class Codes: G10L 2015/226 using non-speech characteristics
• CPC Class Codes: G10L 2015/227 of the speaker; Human-factor methodology
• CPC Class Codes: G10L 2015/228 of application context

Abstract

Provided are a method and an apparatus for speech recognition, and a method and an apparatus for training transformation parameter. A speech recognition apparatus includes an acoustic score calculator configured to use an acoustic model to calculate an acoustic score of a speech input, an acoustic score transformer configured to transform the calculated acoustic score into an acoustic score corresponding to standard pronunciation by using a transformation parameter, and a decoder configured to decode the transformed acoustic score to output a recognition result of the speech input.

Claims

1. A speech recognition apparatus comprising:

an acoustic score calculator configured to use an acoustic model to calculate an acoustic score of a speech input;

an acoustic score transformer configured to transform the calculated acoustic score into an acoustic score corresponding to standard pronunciation by using a transformation parameter; and

a decoder configured to decode the transformed acoustic score to output a recognition result of the speech input.

2. The apparatus of claim 1, further comprising a language score calculator configured to use a language model to output a language score of the speech input,

wherein the decoder is configured to decode the calculated acoustic score based on the language score and output the recognition result.

3. The apparatus of claim 1, further comprising a feedback collector configured to collect verification data regarding the recognition result based on a user feedback.

4. The apparatus of claim 3, further comprising a trainer configured to re-train the transformation parameter based on the acoustic score calculated by the acoustic score calculator and the collected verification data, and update the transformation parameter based on the re-training.

5. The apparatus of claim 3, wherein the trainer updates the transformation parameter by transmitting the calculated acoustic score and the collected verification data to a training apparatus, and by receiving the re-trained transformation parameter from the training apparatus.

6. The apparatus of claim 1, wherein the acoustic score transformer applies the transformation parameter as at least one of a matrix product algorithm, a neural network algorithm, or a linear regression algorithm to the calculated acoustic score.

7. The apparatus of claim 6, wherein the transformation parameter to be applied is determined in advance based on at least one of a user identity, a group identity, an age, a geographic region, a dialect and a gender of a user.

8. A speech recognition method comprising:

using an acoustic model to calculate an acoustic score of a speech input;

transforming the calculated acoustic score into an acoustic score for standard pronunciation by using a transformation parameter; and

decoding the transformed acoustic score to output a recognition result of the speech input.

9. The method of claim 8, further comprising using a language model to output a language score of the speech input,

wherein the decoding of the transformed acoustic score comprises decoding the calculated acoustic score based on the language score.

10. The method of claim 8, further comprising collecting verification data regarding the recognition result based on a user feedback.

11. The method of claim 10, further comprising:

re-training the transformation parameter based on the acoustic score calculated by the acoustic score calculator and the collected verification data; and

updating the transformation parameter based on the re-training.

12. The method of claim 10, further comprising:

transmitting the calculated acoustic score and the collected verification data to a training apparatus;

receiving the re-trained transformation parameter from the training apparatus; and

updating the transformation parameter by using the received transformation parameter.

13. The method of claim 8, wherein the calculating, the transforming and the decoding are performed by one or more processors.

14. A non-transitory computer-readable medium storing instructions that, when executed by a process, causes a processor to perform the method of claim 8.

15. An apparatus for training a transformation parameter, the apparatus comprising:

a training data collector configured to collect, as training data, acoustic score data of actual speech for training, and verification data based on standard pronunciation that corresponds to the actual speech for training;

a trainer configured to train the transformation parameter by using the training data; and

a data transmitter configured to transmit the trained transformation parameter to a speech recognition apparatus.

16. The apparatus of claim 15, wherein:

the training data collector is configured to classify the collected training data according to predetermined criteria; and

the trainer is configured to train the transformation parameter for each classified group to generate the transformation parameter for the each classified group.

17. the apparatus of claim 16, wherein the predetermined criteria comprise at least one of a user identity, a group identity, an age, a geographic region, a dialect and a gender.

18. The apparatus of claim 15, wherein the training data collector further collects, as training data, acoustic score data of a user speech and verification data based on a user feedback, from the speech recognition apparatus.

19. A method of training a transformation parameter, the method comprising:

collecting, as training data, acoustic score data of actual speech and verification data based on standard pronunciation that corresponds to the actual speech for training;

training the transformation parameter by using the training data; and

transmitting the trained transformation parameter to a speech recognition apparatus.

20. The method of claim 19, wherein:

the collecting of the training data comprises classifying the collected training data according to predetermined criteria; and

the training comprises training the transformation parameter for each classified group, and generating the transformation parameter for each classified group.

21. The method of claim 20, wherein the predetermined criteria comprises at least one of a user identity, a group identity, an age, a geographic region, a dialect and a gender.

22. The method of claim 19, wherein the collecting of the training data comprises further collecting, as training data, acoustic score data of a user speech and correct recognition data based on a user feedback, from the speech recognition apparatus.

23. A speech recognition apparatus comprising:

a processor configured to;

calculate a first acoustic score based on a speech input;

convert the first acoustic score into a second acoustic score corresponding to standard pronunciation by using a transformation parameter; and

decode the second acoustic score to recognize the speech input.

24. The apparatus of claim 23, wherein the processor is configured to calculate the first acoustic score with an acoustic model based on standard pronunciation, and

the processor is configured to apply the transformation parameter to the first acoustic score in generating the second acoustic score to account for a pronunciation feature of a user that deviates from the standard pronunciation.

25. The apparatus of claim 23, wherein the processor is configured to use the transformation parameter to adjust for a pronunciation feature of a user that relates to at least one of a user identity, a group identity, an age, a geographic region, a dialect and a gender of the user.

26. An electronic device comprising:

a microphone configured to detect a signal corresponding to the speech input; and

the speech recognition apparatus of claim 23.