Investigation of silicon auditory models and generalization of linear discriminant analysis for improved speech recognition
Andreas G. Andreou, Nagendra Kumar · 1997
Biologically motivated feature extraction algorithms have been found to provide significantly robust performance in speech recognition systems, in the presence of channel and noise degradation, when compared to the standard features such as mel-cepstrum coefficients. However, auditory feature extraction is computationally expensive, and makes these features useless for real-time speech recognition systems. In this thesis, I investigate the use of low power techniques and custom analog VLSI for auditory feature extraction. I first investigated the basilar-membrane model and the hair-cell model chips that were designed by Liu (Liu, 1992). I performed speech recognition experiments to evaluate how well these chips would perform as a front-end to a speech recognizer. Based on the experience gained by these experiments, I propose an alternate architecture that goes beyond the basilar-membrane model, and, using which, auditory features can be computed in real time. These chips have been designed and tested, and consume only a few milliwatts of power as compared to general purpose digital machines that consume several Watts. I have also investigated Linear Discriminant Analysis (LDA) for dimension reduction of auditory features. Researchers have used Fisher-Rao linear discriminant analysis (LDA) to reduce the feature dimension. They model the low-dimensional features obtained from LDA as the outputs of a Markov process with hidden states (HMM). I present a unified framework for the problem of dimension reduction and HMM parameter estimation by modeling the original features with reduced-rank HMM. This re-formulation also leads to a generalization of LDA that is consistent with the heteroscedastic state models used in HMM, and give better performance when tested on a digit recognition task.