SIGNAL BIAS REMOVAL BY MAXIMUM-LIKELIHOOD-ESTIMATION FOR ROBUST TELEPHONE SPEECH RECOGNITION

An acoustical mismatch between the training and the testing conditions of hidden Markov model (HMM)-based speech recognition systems often c auses a severe degradation in the recognition performance. In telephon e speech recognition, for example, undesirable signal components due t o ambient noise and channel distortion, as well as due to different va riations of telephone handsets render the recognizer unusable for real -world applications, This paper presents a signal bias removal (SBR) m ethod based on maximum likelihood estimation for the minimization of t hese undesirable effects, The proposed method is readily applicable in various architectures, i.e., discrete (vector-quantization based), se micontinuous and continuous density HMM, In this paper, the SBR method , integrated into a discrete density HMM, is applied to telephone spee ch recognition where the contamination due to extraneous signal compon ents is assumed to be unknown, To enable real-time implementation, a s equential method for the estimation of the bias is presented, Experime ntal results for speaker-independent connected digit recognition show a reduction in the per digit error rate by up to 41% and 14% during mi smatched and matched training and testing conditions, respectively.