In this thesis, a shared memory based multiprocessor system used as speech recognizer is proposed and developed. It allows up to seven TMS320C30 digital signal processors executed in parallel and had about 233.31 million floating point operations per second in ideal cases. Beside the mentioned application, a CELP speech coder is also ported to the multiprocessor system in order to demonstrate the performance of the system in the other fields of speech processing. Experimental results shown that the proposed multiprocessor system has a better performance than an uniprocessor system. Dynamic time warping (DTW) and hidden Markov model (HMM) are two state-of- the-art speech recognition algorithms. They had been implemented in many successful speech recognition systems. In order to enhance the performance of the algorithms, a stochastic search method called Genetic Algorithm (GA) is presented for nonlinear time-normalization and HMM training. The global searching nature of the GA can achieve better performance than the DTW and the classical HMM training. In addition, the intrinsic parallel characteristic of the GA allows the algorithms executed in the multiprocessor system efficiently. Experimental results shown that better results can be obtained by using GA in the nonlinear time-normalization and HMM training. A review of several speech analysis methods is also described in this work. It includes the discrete Fourier transform, the fast Fourier transform, the filter-bank, the cepstral analysis, the linear predictive analysis, and the pitch detection. This review aimed to find a robust feature extractor for our system, so that a fair comparison between the algorithms can be performed. Finally, cepstral analysis is selected as the feature extractor of the all experiments in this thesis. It is because of the noise resistive properties of the cepstral analysis can reduce the effect of noise on the experimental results.
Author: Chau, Dennis Chak Wai
Source: City University of Hong Kong
Download URL 2: Visit Now