Recognizing Packet Loss Speech Using Client-End and Server-End Mitigation Techniques

• Main Authors: Yi-Jia Huang, 黃奕嘉
• Other Authors: Fu-Rong Jean
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/txgp84

碩士 === 國立臺北科技大學 === 電機工程研究所 === 105 === Due to the rapid evolution of artificial intelligence technology, distributed speech recognition now has becoming a convenient standard human-machine interface on smart mobile devices. Speech feature is extracted at the client-end and is sent through the network channel to the server-end for recognition. Speech feature packets may be lost due to inevitably delay or transmission error over error-prone channels. The purpose of the theses aims at finding ways of reducing the effect of packet loss on speech recognition. In this thesis, we use Mel-frequency cepstral coefficients (MFCC) as the speech feature parameter. Moreover, matrix interleaving at the client-end and matrix de-interleaving at the server-end are applied to shape the burst-like packet loss into uniform-like distribution prior to recognition resulting in less damage. Furthermore, the standard ETSI-duplication or feature interpolation is exploited for reconstructing missing frames. In addition to the full frame rate decoder, speech recognition is also experimented by using the weighted Viterbi algorithm (WVA) and model adaptation (MA) method. The experimental results show that using matrix interleaving and matrix de-interleaving indeed can reduce the length of continuous missing frames, and the higher the matrix interleaving order is, the less the recognition damage should be. As we can see in the experiments, the performance of feature interpolation is slightly better than that of ETSI-duplication. Among all, the model adaptation method achieves the highest recognition rate in average of all SNR conditions.