EUROSPEECH 2003 - INTERSPEECH 2003
This paper examines techniques of discriminative optimization for acoustic model, including both HMM parameters and linear transforms, in the context of HUB5 Mandarin large vocabulary speech recognition task, with the aim to partly solve the problems brought by the sparseness and the highly ambiguous nature of the telephony conversational speech data. Three techniques are studied: MMI training of the HMM acoustic parameters, MMI training of Semi-Tied Covariance Model and MMI Speak Adaptive Training. Descriptions of our recognition system and the algorithms used in our experiments will be detailed, followed by the corresponding results.
Bibliographic reference. Ding, Peng / Chen, Zhenbiao / Hu, Sheng / Zhang, Shuwu / Xu, Bo (2003): "Discriminative optimization of large vocabulary Mandarin conversational speech recognition system", In EUROSPEECH-2003, 1965-1968.