Cross-stream observation dependencies for multi-stream speech recognition
Özgür Çetin, Mari Ostendorf · 2003
This paper extends prior work in multi-stream modeling by introducing cross-stream observation dependencies and a new discriminative criterion for selecting such dependencies. Experimental results combining short-term PLP features with longterm TRAP features show gains associated with a multi-stream model with partial state asynchrony over a baseline HMM. Frame-based analyses show significant discriminant information in the added cross-stream dependencies, but so far there are only small gains in recognition accuracy. 1.