A minimax classification approach with application to robust speech recognition

Abstract

Abstruct-A minimax approach for robust classification of parametric information sources is studied and applied to isolatedword speech recognition based on hidden Markov modeling. The goal is to reduce the sensitivity of speech recognition systems to a possible mismatch between the training and testing conditions. To this end, a generalized likelihood ratio test is developed and shown to be optimal in the sense of achieving the highest asymptotic exponential rate of decay of the error probability for the worst-case mismatch situation. The proposed approach is compared to the standard approach, where no mismatch is assumed, in recognition of noisy speech and in other realistic mismatch situations.

A survey of hybrid ANN/HMM models for automatic speech recognition

“…In addition to the application of the gradient method, Moon and Hwang [84] propose also an inversion technique directly based on the Baum}Welch algorithm, as well as a combined scheme where the HMM inversion is combined with a MINIM-AX-based [82] model parameter adaptation [68]. Parameter adaptation can be accomplished in a batch manner, as well as alternating it with an inversion step, in an iterative fashion.…”

A Bayesian predictive classification approach to robust speech recognition

“…This provides the BPC approach a way to be different from the conventional plug-in MAP decoder. As for the relation between our BPC approach and other robust decision approaches such as the approximate Bayesian decision rule in [13], the minimax decision rule in [14], readers are referred to [4] for a detailed discussion.…”

“…By further replacing V in (14) withŨ ; we can evaluate the approximate predictive pdf in (14) and perform BPC-based recognition. The resulting BPC rule is thus named as the quasi-Bayes predictive classification, or QBPC, rule.…”

“…In [14], the power spectral density (PSD) of a block of speech signal (one speech frame of short-time spectral analysis), S(! ); is assumed to be represented by a rational function of e j!…”

“…Although in many practical speech recognition systems, some empirical cepstral representations such as MFCC (mel-frequency cepstral coefficients) and LPCC (linear predictive coding cepstrum) are actually used, the above result still approximately holds for these speech representations. This fact motivates the authors of [14] to adopt a uniform distribution for mean vectors of CDHMM in an uncertainty neighborhood of λ as follows:…”