Minimax Classification with Parametric Neighborhoods for Noisy Speech Recognition

In this paper we derive upper and lower bounds on the mean of speech signals corrupted by additive noise. The bounds are derived in the log spectral domain. Approximate bounds on the first and second order time derivatives are also developed. It is then shown how to transform these bounds to the MFCC domain to be used by conventional cepstrum-based speech recognizers. The proposed bounds define the mismatch neighborhood for minimax classification. Speech recognition experiments, using artificially added noise, and a real-life mismatch scenario, illustrate that this parametric neighborhood works quite well in practice. We also believe that the proposed bounds will find various applications in noisy speech recognition.