Generating topic-specific language models

• US 9,892,730 B2
• Filed: 07/01/2009
• Issued: 02/13/2018
• Est. Priority Date: 07/01/2009
• Status: Active Grant

• Current Assignee: TiVo Corporation (Adeia Inc.)
• Original Assignee: Comcast Interactive Media LLC (Comcast Corporation)
• Inventors: Houghton, David F., Murray, Seth Michael, Simon, Sibley Verbeck
• Primary Examiner(s): BAKER, MATTHEW H
• Application Number:
  US12/496,081
• Publication Number:
  US 20110004462A1
• Time in Patent Office: 3,149 Days
• Field of Search: 704 1- 10
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams

Abstract

Speech recognition may be improved by generating and using a topic specific language model. A topic specific language model may be created by performing an initial pass on an audio signal using a generic or basis language model. A speech recognition device may then determine topics relating to the audio signal based on the words identified in the initial pass and retrieve a corpus of text relating to those topics. Using the retrieved corpus of text, the speech recognition device may create a topic specific language model. In one example, the speech recognition device may adapt or otherwise modify the generic language model based on the retrieved corpus of text.

Claims

1. A method comprising:

performing, by a computing device and using a first language model, a first speech recognition process on an audio signal;

determining, by the computing device and based on the first speech recognition process, a plurality of topics associated with the audio signal;

determining, by the computing device and based on the first speech recognition process, a respective significance, for each of the plurality of topics, based on a respective quantity of terms, in the audio signal, associated with each of the plurality of topics;

determining, by the computing device and based on the respective significance for each of the plurality of topics, a respective term threshold;

causing, for each of the plurality of topics, a respective set of one or more searches such that a respective quantity of terms identified by the respective set of one or more searches satisfies the respective term threshold for the topic;

determining, by the computing device and based on the terms identified by the searches, a second language model; and

performing, by the computing device and using the second language model, a second speech recognition process on the audio signal.

2. The method of claim 1, wherein the determining the plurality of topics associated with the audio signal comprises:

determining, using the first language model, one or more spoken terms including a first term corresponding to a first topic associated with the audio signal; and

determining one or more of;

a frequency of the first term corresponding to the first topic, or

whether first the term corresponding to the first topic appears in a list of stop words.

3. The method of claim 1, further comprising:

determining whether a quantity of terms returned by a first search satisfies the respective term threshold for a first topic of the plurality of topics;

in response to determining that the quantity of terms returned by the first search does not satisfy the respective term threshold for the first topic, conducting a second search associated with the first topic; and

extracting text from a second plurality of search results corresponding to the second search.

4. The method of claim 1, wherein the determining the second language model comprises determining a probability of a first term following a second term.

10. The method of claim 1, further comprising:

determining a plurality of most-frequently-used terms in the audio signal,

wherein the determining the plurality of topics comprises determining a first topic corresponding to at least one term of the plurality of most-frequently-used terms in the audio signal.

11. The method of claim 10, further comprising:

determining a second topic, associated with the audio signal, based on a different term of the plurality of most-frequently-used terms in the audio signal.

12. The method of claim 1, wherein the first language model is a generic language model.

13. The method of claim 1, wherein the determining, for each topic of the plurality of topics, the respective term threshold further comprises:

dividing a total number of terms needed to generate the second language model by a total number of topics in the plurality of topics.

14. The method of claim 1, further comprising:

determining that an audio file is located on a web page, wherein the audio file is associated with the audio signal; and

extracting data from the web page, independent of the audio file, to determine a topic of the web page, wherein the determining the plurality of topics associated with the audio signal is based on the topic of the web page.

5. A method comprising:

determining, by a computing device and via a first speech recognition process, a first topic and a second topic associated with an audio signal, wherein the first speech recognition process uses an initial language model;

determining, by the computing device, a significance of the first topic based on a first quantity of terms, in the audio signal, identified as being associated with the first topic via the first speech recognition process;

determining, by the computing device, a significance of the second topic based on a second quantity of terms, in the audio signal, identified as being associated with the second topic via the first speech recognition process;

receiving, by the computing device and in response to a first search associated with the first topic, a first plurality of terms related to at least the first topic, wherein a quantity of the first plurality of terms satisfies a first threshold number of terms that are based on the significance of the first topic;

causing, by the computing device and based on the first plurality of terms, modification of the initial language model; and

performing, by the computing device and using the modified initial language model, a second speech recognition process on the audio signal.

6. The method of claim 5, wherein the determining the second topic associated with the audio signal is based on an identification of at least one stop word in the audio signal and further based on metadata for the audio signal.

7. The method of claim 6, further comprising:

accessing a word list comprising a plurality of stop words;

determining that a word in the audio signal is on the word list; and

determining the second topic associated with the audio signal based on a different word than the word in the audio signal that is on the word list.

8. The method of claim 7, wherein the word list is a topic-specific stop word list.

9. The method of claim 6, further comprising:

determining that the at least one stop word in the audio signal is one of a plurality of pre-designated stop words.

15. The method of claim 5, further comprising:

receiving, by the computing device and in response to a second search associated with the second topic, a second plurality of terms related to the second topic, wherein a quantity of the second plurality of terms satisfies a second threshold number of terms that are based on the significance of the second topic, the method further comprising;

causing modification of the initial language model based at least in part on the second plurality of terms.

16. A method comprising:

performing, by a computing device and using a first language model, a first speech recognition process on an input signal;

determining, based on the first speech recognition process, a plurality of topics associated with the input signal;

determining, for each topic of the plurality of topics;

a respective significance based on a respective quantity of terms, in the input signal, associated with each of the plurality of topics; and

a respective term threshold based on the respective significance;

causing, for each of the plurality of topics and using words recognized by the first speech recognition process, one or more searches such that a quantity of terms identified by the one or more searches satisfies the respective term threshold for the topic;

determining a corpus of terms by combining the terms returned by the one or more searches conducted for each of the plurality of topics;

determining, based on the corpus of terms, a second language model; and

performing, by the computing device and using the second language model, a second speech recognition process on the input signal.

17. The method of claim 16, wherein the determining the respective significance for a first topic of the plurality of topics comprises determining a number of words or phrases, identified by the first speech recognition process, as being associated with the first topic.

18. The method of claim 16, wherein the causing the one or more searches for a first topic of the plurality of topics comprises iteratively conducting a plurality of searches until a total quantity of terms identified by the iteratively conducted searches satisfies the respective term threshold for the first topic.

19. The method of claim 16, wherein the determining the second language model comprises causing, in the first language model, modification of a probability of two terms appearing consecutively.

20. The method of claim 16, wherein the causing the one or more searches for a first topic of the plurality of topics comprises retrieving, from a keyword table, a plurality of keywords previously associated with the first topic.