A maximum likelihood equalization technique for robust speech recognition in adverse environments

In this paper, we study the problem of robust speech recognition in adverse environments. We focus our attention to the following two types of distortions: 1) the additive noise distortion, 2) the channel mismatch distortion. The maximum likelihood (ML) equalization technique is used to compensate for these distortions. Performance of the ML technique is compared with the following channel equalization techniques: the global mean subtraction (GMS) technique, the local mean subtraction (LMS) technique, the finite impulse response (FIR) highpass filtering technique, the infinite impulse response (IIR) highpass filtering technique, the RASTA (bandpass) filtering technique, and the masking-based filtering technique. These techniques have been recently proposed in the literature and are computationally much simpler than the ML equalization technique. It is shown that the ML equalization technique does not offer any significant advantage over the other channel equalization techniques in terms of recognition performance.