Path-dependent kalman estimation of a cepstral bias

An acoustic mismatch between a given utterance and a model degrades the performance of the speech recognition process. We choose to model speech by Hidden Markov Models (HMMs) in the cepstrum domain and the mismatch by an additive bias. To track the variations of this bias, we explicitly model the way in which the bias can vary by a state equation. We derive a frame-synchronous estimator of this bias based on Kalman recursions. We use this estimator to compensate for the mismatch in the recognition process. Finally, we report recognition experiments carried out over both public switched telephone network (PSTN) and cellular telephone network to show the efficiency of the method in a real context.