METHOD FOR ESTIMATING LANGUAGE MODEL WEIGHT AND SYSTEM FOR THE SAME

• US 20120150539A1
• Filed: 12/13/2011
• Published: 06/14/2012
• Est. Priority Date: 12/14/2010
• Status: Active Grant

• Current Assignee: Electronics and Telecommunications Research Institute
• Original Assignee: Electronics and Telecommunications Research Institute
• Inventors: Jeon, Hyung Bae, Lee, Yun Keun, Chung, Eui Sok, Kim, Jong Jin, Chung, Hoon, Park, Jeon Gue, Jung, Ho Young, Park, Ki Young, Lee, Sung Joo, Kang, Jeom Ja, Kang, Byung Ok, Song, Hwa Jeon
• Granted Patent: US 8,666,739 B2
• Time in Patent Office: Days
• US Class Current: 704/236
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/187 Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams

Abstract

Method of the present invention may include receiving speech feature vector converted from speech signal, performing first search by applying first language model to the received speech feature vector, and outputting word lattice and first acoustic score of the word lattice as continuous speech recognition result, outputting second acoustic score as phoneme recognition result by applying an acoustic model to the speech feature vector, comparing the first acoustic score of the continuous speech recognition result with the second acoustic score of the phoneme recognition result, outputting first language model weight when the first coustic score of the continuous speech recognition result is better than the second acoustic score of the phoneme recognition result and performing a second search by applying a second language model weight, which is the same as the output first language model, to the word lattice.

Claims

1. A system for estimating a language model weight, the system comprising:

a first search unit which receives a speech feature vector converted from a speech signal, performs a first search by applying a first language model to the received speech feature vector, and outputs a word lattice and a first acoustic score of the word lattice as a continuous speech recognition result;

a phoneme recognition unit which outputs a second acoustic score as a phoneme recognition result by applying an acoustic model to the speech feature vector;

a weight estimation unit which compares the first acoustic score of the continuous speech recognition result of the first search unit with the second acoustic score of the phoneme recognition result of the phoneme recognition unit and, when the first acoustic score of the continuous speech recognition result is better than the second acoustic score of the phoneme recognition result, outputs a first language model weight; and

a second search unit which performs a second search by applying a second language model weight, which is the same as the output first language model weight, to the word lattice and outputs a final recognition result.

2. The system of claim 1, wherein when the first acoustic score of the continuous speech recognition result is worse than the second acoustic score of the phoneme recognition result, the weight estimation unit outputs a weight which is lower than the first language model weight.

3. The system of claim 2, wherein the second search unit performs the second search by applying a second language model weight which is lower than the first language model weight.

4. The system of claim 1, wherein when the first acoustic score of the continuous speech recognition result is neither better nor worse than the second acoustic score of the phoneme recognition result, the weight estimation unit outputs a weight corresponding to an intermediate value between the first language model weight and the second language model weight.

5. The system of claim 4, wherein the second search unit performs the second search by applying the weight corresponding to the intermediate value between the first language model weight and the second language model weight.

6. The system of claim 1, wherein the word lattice is obtained by defining a plurality of word combinations searched by the first search as information about connections between words.

7. The system of claim 1, wherein the word lattice comprises information used to evaluate the continuous speech recognition result of the first search unit.

8. A method for estimating a language model weight, the method comprising:

receiving a speech feature vector converted from a speech signal, performing a first search by applying a first language model to the received speech feature vector, and outputting a word lattice and a first acoustic score of the word lattice as a continuous speech recognition result;

outputting a second acoustic score as a phoneme recognition result by applying an acoustic model to the speech feature vector;

comparing the first acoustic score of the continuous speech recognition result with the second acoustic score of the phoneme recognition result;

outputting a first language model weight when the first acoustic score of the continuous speech recognition result is better than the second acoustic score of the phoneme recognition result; and

performing a second search by applying a second language model weight, which is the same as the output first language model, to the word lattice.

9. The method of claim 8, wherein in the outputting of the weight, when the first acoustic score of the continuous speech recognition result is worse than the second acoustic score of the phoneme recognition result, a weight which is lower than the first language model weight is output.

10. The method of claim 9, wherein in the performing of the second search, a second language model weight which is lower than the first language model weight is applied.

11. The method of claim 8, wherein in the outputting of the weight, when the first acoustic score of the continuous speech recognition result is neither better nor worse than the second acoustic score of the phoneme recognition result, a weight corresponding to an intermediate value between the first language model weight and the second language model weight is output.

12. The method of claim 11, wherein in the performing the second search, the weight corresponding to the intermediate value between the first language model weight and the second language model weight is applied.

13. The method of claim 8, wherein the word lattice is obtained by defining a plurality of word combinations searched by the first search as information about connections between words.

14. The method of claim 8, wherein the word lattice comprises information used to evaluate the continuous speech recognition result of the first search unit.