TY - JOUR
T1 - Automatic query generation and query relevance measurement for unsupervised language model adaptation of speech recognition
AU - Ito, Akinori
AU - Kajiura, Yasutomo
AU - Suzuki, Motoyuki
AU - Makino, Shozo
PY - 2009
Y1 - 2009
N2 - We are developing a method of Web-based unsupervised language model adaptation for recognition of spoken documents. The proposed method chooses keywords from the preliminary recognition result and retrieves Web documents using the chosen keywords. A problem is that the selected keywords tend to contain misrecognized words. The proposed method introduces two new ideas for avoiding the effects of keywords derived from misrecognized words. The first idea is to compose multiple queries from selected keyword candidates so that the misrecognized words and correct words do not fall into one query. The second idea is that the number of Web documents downloaded for each query is determined according to the "query relevance". Combining these two ideas, we can alleviate bad effect of misrecognized keywords by decreasing the number of downloaded Web documents from queries that contain misrecognized keywords. Finally, we examine a method of determining the number of iterative adaptations based on the recognition likelihood. Experiments have shown that the proposed stopping criterion can determine almost the optimum number of iterations. In the final experiment, the word accuracy without adaptation (55.29) was improved to 60.38, which was 1.13 point better than the result of the conventional unsupervised adaptation method (59.25).
AB - We are developing a method of Web-based unsupervised language model adaptation for recognition of spoken documents. The proposed method chooses keywords from the preliminary recognition result and retrieves Web documents using the chosen keywords. A problem is that the selected keywords tend to contain misrecognized words. The proposed method introduces two new ideas for avoiding the effects of keywords derived from misrecognized words. The first idea is to compose multiple queries from selected keyword candidates so that the misrecognized words and correct words do not fall into one query. The second idea is that the number of Web documents downloaded for each query is determined according to the "query relevance". Combining these two ideas, we can alleviate bad effect of misrecognized keywords by decreasing the number of downloaded Web documents from queries that contain misrecognized keywords. Finally, we examine a method of determining the number of iterative adaptations based on the recognition likelihood. Experiments have shown that the proposed stopping criterion can determine almost the optimum number of iterations. In the final experiment, the word accuracy without adaptation (55.29) was improved to 60.38, which was 1.13 point better than the result of the conventional unsupervised adaptation method (59.25).
UR - http://www.scopus.com/inward/record.url?scp=77649264779&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=77649264779&partnerID=8YFLogxK
U2 - 10.1155/2009/140575
DO - 10.1155/2009/140575
M3 - Article
AN - SCOPUS:77649264779
VL - 2009
JO - Eurasip Journal on Audio, Speech, and Music Processing
JF - Eurasip Journal on Audio, Speech, and Music Processing
SN - 1687-4714
M1 - 140575
ER -