Investigation of silicon auditory models and generalization of linear discriminant analysis for improved speech recognition

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