Speech Feature Compression Using SVD and Optimal SQ in Distributed Speech Recognition Systems

• Main Authors: Ting-En Wu, 吳廷恩
• Other Authors: 簡福榮
• Format: Others
• Online Access: http://ndltd.ncl.edu.tw/handle/79sb94

碩士 === 國立臺北科技大學 === 電機工程研究所 === 105 === In distributed speech recognition (DSR) systems, speech recognition is split into two major components, feature extraction is performed at the client-end, then the extracted speech feature is sent and transmitted through communication channels to the back-end server for recognition. In this thesis, two speech features are investigated including mel-frequency cepstral coefficients (MFCC) and perceptual linear predictive cepstral coefficients (PLPCC). In order to achieve low bit-rate transmission, we employ singular value decomposition (SVD) and optimal scalar quantization (OSQ) for feature compression. In the proposed SVD_OSQ scheme, the speech features are grouped into blocks of nine or ten successive frames without overlapping. The half frame rate (1/2 frame rate) and one-third frame rate (1/3 frame rate) data using feature interpolation reconstruction (FIR) at the back-end are experimented as the baseline systems for comparison. The experimental results show that the recognition performance of the proposed SVD_OSQ scheme is found to be superior to that of the FIR scheme when using MFCC feature. Furthermore, for multi-condition training model the recognition performance of using PLPCC feature is better than that of using MFCC feature.