Unsupervised training method, training apparatus, and training program for an N-gram language model based upon recognition reliability

Patent metadata

Field	Value
Publication number	US-9536518-B2
Application number	US-201514643316-A
Country	US
Kind code	B2
Filing date	Mar 10, 2015
Priority date	Mar 27, 2014
Publication date	Jan 3, 2017
Grant date	Jan 3, 2017

Abstract

A computer-based, unsupervised training method for an N-gram language model includes reading, by a computer, recognition results obtained as a result of speech recognition of speech data; acquiring, by the computer, a reliability for each of the read recognition results; referring, by the computer, to the recognition result and the acquired reliability to select an N-gram entry; and training, by the computer, the N-gram language model about selected one of more of the N-gram entries using all recognition results.

First claim

The invention claimed is:

1. An unsupervised training system for an N-gram language model, comprising:
a processor configured to:
read recognition results obtained as a result of speech recognition of speech data;
acquire a reliability for each of the read recognition results;
refer to each recognition result's acquired reliability to select a subset of one or more N-gram entries based upon their respective reliabilities; and
train an N-gram language model for one of more entries of the subset of N-gram entries using all recognition results, wherein the processing device is further configured to select from a first corpus, a second corpus, and a third corpus, each of the N-gram entries, whose sum of a first number of appearances in the first corpus as a set of all the recognition results, a second number of appearances in a second corpus as a subset of the recognition results with the reliability higher than or equal to a predetermined threshold value, and a third number of appearances in the third corpus as a baseline of the N-gram language model exceeds a predetermined number of times, where each of the first number of appearances, the second number of appearances, and the third number of appearances is given a different weight, respectively.

2. The system of claim 1 , wherein each of the weights respectively given to each of the first number of appearances, the second number of appearances, and the third number of appearances is estimated in advance by an EM algorithm using a language model estimated from each of subsets of the first corpus, the second corpus, and the third corpus.

3. The system of claim 1 , wherein the processing device is further configured to train the selected one or more N-gram entries using all the recognition results and adding, to a base N-gram language model, the one or more N-gram entries and probabilities obtained as a result of the training.

4. The system of claim 1 , wherein the acquired recognition results of the speech data are recognition results as a result of automatic speech recognition in a cloud type speech recognition system or a server type speech recognition system.

5. The system of claim 1 , wherein a posterior probability of a text unit obtained upon speech recognition of the speech data is used as the reliability in the acquiring.

6. A non-transitory, computer readable storage medium having instructions stored thereon that, when executed by a computer, implement a training method for an N-gram language model, the method comprising:
reading recognition results obtained as a result of speech recognition of speech data;
acquiring a reliability for each of the read recognition results;
referring to each recognition result's acquired reliability to select a subset of one or more N-gram entries based upon their respective reliabilities; and
training the N-gram language model for one of more entries of the subset of N-gram entries using all recognition results, wherein the processing device is further configured to select each of the N-gram entries, whose sum of a first number of appearances in a first corpus as a set of all the recognition results and a second number of appearances in a second corpus as a subset of the recognition results with the reliability higher than or equal to a predetermined threshold value exceeds a predetermined number of times.

7. The computer readable storage medium of claim 6 , wherein the method further comprises:
training the selected one or more N-gram entries using all the recognition results and adding, to a base N-gram language model, the one or more N-gram entries and probabilities obtained as a result of the training.

8. The computer readable storage medium of claim 6 , wherein the acquired recognition results of the speech data are recognition results as a result of automatic speech recognition in a cloud type speech recognition system or a server type speech recognition system.

9. The computer readable storage medium of claim 6 , wherein a posterior probability of a text unit obtained upon speech recognition of the speech data is used as the reliability in the acquiring.

10. An unsupervised training system for an N-gram language model, comprising:
a processor configured to:
read recognition results obtained as a result of speech recognition of speech data;
acquire a reliability for each of the read recognition results;
refer to each recognition result's acquired reliability to select a subset of one or more N-gram entries based upon their respective reliabilities; and
train an N-gram language model for one of more entries of the subset of N-gram entries using all recognition results, wherein the processor is further configured to select each of the N-gram entries, whose sum of a first number of appearances in a first corpus as a set of all the recognition results and a second number of appearances in a second corpus as a subset of the recognition results with the reliability higher than or equal to a predetermined threshold value exceeds a predetermined number of times.

11. The system of claim 10 , wherein the processor is further configured to train the selected one or more N-gram entries using all the recognition results and adding, to a base N-gram language model, the one or more N-gram entries and probabilities obtained as a result of the training.

12. The system of claim 10 , wherein the acquired recognition results of the speech data are recognition results as a result of automatic speech recognition in a cloud type speech recognition system or a server type speech recognition system.

13. The system of claim 10 , wherein a posterior probability of a text unit obtained upon speech recognition of the speech data is used as the reliability in the acquiring.

Assignees

• IBM

Inventors

• Itoh Nobuyasu
• Kurata Gakuto
• Nishimura Masafumi

Classifications

• G10L15/30

  Distributed recognition, e.g. in client-server systems, for mobile phones or network applications · CPC title

• G10L15/197

  Probabilistic grammars, e.g. word n-grams · CPC title

• G10L2015/0635

  updating or merging of old and new templates; Mean values; Weighting · CPC title

• G10L15/18

  using natural language modelling · CPC title

• G10L15/063 Primary

  Training · CPC title