Recognition of Speech from Live Sports Coverage Using Acoustic and Language Model Adaptation

摘要

In this paper we use large vocabulary continuous speech recognition to detect highlight scenes in live radio coverage of baseball games and extract information for indexing (keywords). To make the speech recognition unit more robust we perform MLLR and MAP adaptation of the acoustic models proposing a two-level adaptation procedure using both supervised and unsupervised adaptation. By performing acoustic model adaptation, we are able to adapt to the announcer's speaking style and the recording environment, achieving an improvement of 28% in word accuracy over the baseline. Regarding the language model, we performed adaptation via language model fusion, created classes for players' and commentators' names, and adjusted the pronunciation dictionary, resulting in an improvement of 13% in word accuracy over the baseline. By combined adaptation of both models we achieved an improvement of 38% in word accuracy.