| Summary: | 碩士 === 國立成功大學 === 電機工程學系碩博士班 === 97 === In this thesis, the fixed-point arithmetic design for embedded text-independent speaker recognition system is proposed. The goal is to increase the convenience of life by portable devices. The Linear Prediction Cepstrum Coefficients (LPCC) is used for feature extraction. The Multi-Class Support Vector Machines (SVMs) algorithm is adopted for speaker model training and speaker classification.
In order to overcome the time-consuming problem in speaker model training for portable device, the modification of training and classification algorithm is essential. After analyzing the entire algorithm of the system, the processes for performing LPCC extraction and SVMs are found to have a large burden in computation. Therefore, fixed-point implementation is proposed in this thesis. The experimental results show that there is great improvement in training time. Moreover, there are 10.88 times improvement in training time, and 6.51 times improvement in classification time comparing with floating-point implementation. Although the speaker recognition accuracy is decreased due to the truncation error between fixed-point and floating-point design, the speaker recognition accuracy rate can still reach 93.61% in our proposed work.
|
|---|
