| Summary: | 碩士 === 國立成功大學 === 電機工程學系碩博士班 === 96 === A novel architecture for ubiquitous and robust of text-independent speaker recognition based on SVM approach is proposed. In this architecture, multiple far-field microphones of configuration is adopted to receive the pervasive speech signals, and the distance effect between speaker and microphone is supposed to be ignored. Then the multi-channel speech signals are added together through a mixer. In a ubiquitous computing environment, the received speech signal is usually heavily corrupted by background noises. An SNR-aware subspace speech of enhancement approach is used as a pre-processing to enhance the mixed informational signal as well as suppressing the noise. Considering the text-independent speaker recognition, this proposed work applies multi-class support vectors machine (SVM) instead of using conventional Gaussian mixture models (GMMs). In our experiments, the speaker recognition rate up to 97.2% with the proposed ubiquitous architecture of speaker recognition system.
Additionally, we proposed a hardware realization of speaker identification system based on sequential minimal optimization (SMO) algorithm of SVM. We also proposed more efficient method of cache table utilization, and intend to save more then one half of cache table space as well as to reduce processing time of kernel function. Moreover, the heuristics selection method of SMO algorithm is implemented into hardware design to reduce the training time. In our experiments, the training time can reduce 2.17 times less than non-use of heuristics selection method on PC. And our finding shows that the identification ratio up to 92.5% of accuracy and reduced 53% of training time in hardware implementation.
|
|---|
