5th International Conference on Spoken Language Processing
We propose a new approach to improve the performance of speech recognizers by utilizing acoustic-phonetic knowledge sources. We use the unvoiced, voiced. and silence (UVS) group information of the input speech signal in the conventional speech recognizer. We extract the UVS information by, using a recurrent neural network (RNN). generate a rule-based score, and then add the score representing the INS information to the conventional spectral feature-driven score in the search module. Experimental results showed that the approach reduces 9% of errors in a 5000- word Korean spontaneous speech recognition domain.
Bibliographic reference. Suh, Youngjoo / Hwang, Kyuwoong / Kwon, Oh-Wook / Park, Jun (1998): "Improving speech recognizer by broader acoustic-phonetic group classification", In ICSLP-1998, paper 0638.