Method and system for automatic domain adaptation in speech recognition applications
View Patent ↗A system and method for adapting a language model to a specific environment by receiving interactions captured the specific environment, generating a collection of documents from documents retrieved from external resources, detecting in the collection of documents terms related to the environment that are not included in an initial language model and adapting the initial language model to include the terms detected.
1. A method for adapting a language model to a specific environment, the method comprising:
receiving, by a processor, a plurality of interactions captured in the specific environment;
retrieving, by the processor, a plurality of external documents based on a plurality of query expressions, wherein query expressions are generated by applying a topic detection algorithm on textual transcripts in order to detect the different topics discussed in the plurality of interactions;
generating, by the processor, an external corpus by selecting documents from the documents retrieved;
detecting in the external corpus terms related to the environment that are not included in an initial language model; and
adapting, by the processor, the initial language model to include a plurality of the terms detected.
2. The method of claim 1 , comprising automatically transcribing each of the plurality of the interactions into a textual transcript based on the initial language model.
3. The method of claim 1 , wherein generating the external corpus comprises selecting a plurality of documents from the documents retrieved based on semantic similarity between the documents retrieved and the plurality of interactions.
4. The method of claim 1 , wherein detecting terms related to the environment comprises performing topic modeling on a corpora comprising the external corpus.
5. The method of claim 4 , wherein detecting terms related to the environment comprises performing validation of key phrases from the corpora.
6. The method of claim 5 , wherein detecting terms related to the environment comprises performing phonetic validation of the key phrases from the corpora.
7. The method of claim 5 , wherein adapting the initial language model comprises modifying the initial language model such that a speech recognition model is to recognize the terms detected.
8. The method of claim 1 , comprising:
testing the adapted language model; and
storing the adapted model if the adapted model is expected to improve the word error rate; and
performing the adapting operation again if the adapted model is not expected to improve the word error rate of the initial language model.
9. A system for adapting a language model to a specific environment, the system comprising:
a memory;
a recorder to capture a plurality of interactions in the specific environment; and
a computer:
to retrieve a plurality of external documents based on a plurality of query expressions, wherein query expressions are generated by applying a topic detection algorithm on textual transcripts in order to detect the different topics discussed in the plurality of interactions,
to generate an external corpus by selecting a plurality of documents from the documents retrieved,
to detect in the external corpus terms related to the environment that are not included in an initial language model, and
to adapt the initial language model to include a plurality of the terms detected.
10. The system of claim 9 , wherein the computer is to automatically transcribe each of the plurality of the interactions into a plurality of textual transcripts based on the initial language model.
11. The system of claim 9 , wherein the computer is to generate the external corpus by selecting a plurality of documents from the documents retrieved based on semantic similarity between the documents retrieved and the plurality of interactions.
12. The system of claim 9 , wherein the computer is to detect the terms related to the environment by performing topic modeling on a corpora comprising the external corpus.
13. The method of claim 12 , wherein the computer is to detect the terms related to the environment by performing validation of key phrases from the corpora.
14. The system of claim 13 , wherein the computer is to detect the terms related to the environment comprises performing phonetic validation of the key phrases from the corpora.
15. The system of claim 13 , wherein the computer is to modify the initial language model such that a speech recognition model is to recognize the terms detected.
16. A method for adapting a language model, the method comprising:
retrieving, by a processor, a plurality of documents based on a plurality of query expressions, wherein query expressions are generated by applying a topic detection algorithm on textual transcripts in order to detect the different topics discussed in a plurality of interactions;
identifying, based on the plurality of documents, terms associated with an environment, wherein the identified terms are not included in a language model; and
augmenting the language model to the identified terms detected.
17. The method of claim 16 , wherein the plurality of documents are retrieved from external resources.
18. The method of claim 16 , further comprising:
receiving a plurality of interactions captured in the environment; and
automatically transcribing each of the plurality of the interactions into a plurality of textual transcripts based on the language model.