CHANNEL ESTIMATION FOR REFERENCE MODEL ADAPTATION IN TELEPHONE SPEECH RECOGNITION

This paper deals with a channel estimation problem in speech recognition over telephone networks. In this study, we propose a channel estimation method to adapt the reference models so that the reference models can match the testing environment. For a telephone speech, its channel cepstral vector is estimated by the maximum a posteriori criterion. The reference models are then adapted to the testing environment by merging the estimated channel cepstral vectors. The telephone speech is recognized by using the adapted reference models. Experiments show that the proposed method can well estimate the telephone channel spectra and is successfully applied for telephone speech recognition.