Sampling Training Data for an Automatic Speech Recognition System Based on a Benchmark Classification Distribution

• US 20130289989A1
• Filed: 01/18/2013
• Published: 10/31/2013
• Est. Priority Date: 04/26/2012
• Status: Active Grant

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: Google Inc. (Alphabet Inc.)
• Inventors: Nakajima, Kaisuke, Biadsy, Fadi, Moreno Mengibar, Pedro J., Bikel, Daniel Martin
• Granted Patent: US 9,202,461 B2
• Time in Patent Office: Days
• US Class Current: 704/243
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models

Abstract

A set of benchmark text strings may be classified to provide a set of benchmark classifications. The benchmark text strings in the set may correspond to a benchmark corpus of benchmark utterances in a particular language. A benchmark classification distribution of the set of benchmark classifications may be determined. A respective classification for each text string in a corpus of text strings may also be determined. Text strings from the corpus of text strings may be sampled to form a training corpus of training text strings such that the classifications of the training text strings have a training text string classification distribution that is based on the benchmark classification distribution. The training corpus of training text strings may be used to train an automatic speech recognition (ASR) system.

Claims

1. A method comprising:

obtaining a benchmark classification distribution;

selecting, by a computing device, training text strings, wherein the training text strings are associated with respective classifications, and wherein the training text strings are selected such that the respective classifications of the selected training text strings are in proportion to the benchmark classification distribution; and

training an automatic speech recognition (ASR) system using the training text strings.

2. The method of claim 1, wherein the benchmark classification distribution is a distribution of topic classifications.

3. The method of claim 1, wherein obtaining the benchmark classification distribution comprises:

transcribing benchmark utterances to respective benchmark text strings; and

determining the benchmark classification distribution from the benchmark text strings.

4. The method of claim 3, wherein the benchmark utterances were made by users in a category of users, and wherein the ASR system is configured to transcribe new utterances made by users in the category of users.

5. The method of claim 4, wherein the benchmark utterances were made by a single user, and wherein the ASR system is configured to transcribe new utterances made by the single user.

6. The method of claim 4, wherein the benchmark utterances were made by users with a particular dialect, and wherein the ASR system is configured to transcribe new utterances made by users with the particular dialect.

7. The method of claim 4, wherein the benchmark utterances were made by users from a particular geographic location, and wherein the ASR system is configured to transcribe new utterances made by users from the particular geographic location.

8. The method of claim 3, wherein there are fewer benchmark text strings than training text strings.

9. The method of claim 1, wherein the training text strings were transcribed from respective utterances.

10. The method of claim 1, wherein the classifications of the training text strings have a training text string classification distribution, and wherein the training text strings are selected such that the training text string classification distribution is substantially similar to a benchmark classification distribution.

11. The method of claim 1, wherein training the ASR system comprises:

training a language model of the ASR system with a combination the training text strings and an additional corpus of text strings that were transcribed from utterances made by users of the ASR system.

12. An article of manufacture including a non-transitory computer-readable medium, having stored thereon program instructions that, upon execution by a computing device, cause the computing device to perform operations comprising:

obtaining a benchmark classification distribution;

selecting training text strings, wherein the training text strings are associated with respective classifications, and wherein the training text strings are selected such that the respective classifications of the selected training text strings are in proportion to the benchmark classification distribution; and

training an automatic speech recognition (ASR) system using the training text strings.

13. The article of manufacture of claim 12, wherein the benchmark classification distribution is a distribution of topic classifications.

14. The article of manufacture of claim 12, wherein obtaining the benchmark classification distribution comprises:

transcribing benchmark utterances to respective benchmark text strings; and

determining the benchmark classification distribution from the benchmark text strings.

15. The article of manufacture of claim 12, wherein the classifications of the training text strings have a training text string classification distribution, and wherein the training text strings are selected such that the training text string classification distribution is substantially similar to a benchmark classification distribution.

16. The article of manufacture of claim 12, wherein training the ASR system comprises:

training a language model of the ASR system with a combination the training text strings and an additional corpus of text strings that were transcribed from utterances made by users of the ASR system.

17. A computing system comprising:

at least one processor;

data storage; and

program instructions in the data storage that, upon execution by the at least one processor, cause the computing system to perform operations comprising;

obtaining a benchmark classification distribution;

selecting training text strings, wherein the training text strings are associated with respective classifications, and wherein the training text strings are selected such that the respective classifications of the selected training text strings are in proportion to the benchmark classification distribution; and

training an automatic speech recognition (ASR) system using the training text strings.

18. The computing system of claim 17, wherein the benchmark classification distribution is a distribution of topic classifications.

19. The computing system of claim 17, wherein obtaining the benchmark classification distribution comprises:

transcribing benchmark utterances to respective benchmark text strings; and

determining the benchmark classification distribution from the benchmark text strings.

20. The computing system of claim 17, wherein the classifications of the training text strings have a training text string classification distribution, and wherein the training text strings are selected such that the training text string classification distribution is substantially similar to a benchmark classification distribution.