الفهرس | Only 14 pages are availabe for public view |
Abstract Ahmed Ismail Ahmed Ismail. Devising a Very Low Bit-Rate VoIP System Using an Arabic Segmental Vocoder. Master of Science dissertation, Ain Shams University, Faculty of Engineering, 2008. Voice over Internet Protocol (VoJP) has been an interesting topic in the recent years. It is cheaper and well positioned for providing a multitude of different applications other than the conventional two peers’ handset conversation; low bandwidth conversations, multi-conferencing, audio commentary archival and serving over the web, encryption, and internet messaging. This thesis describes research investigating the use of segmental vocoders to provide a very low bit-rate codec.for use in VoIP applications. The research was done with emphasis on the Arabic language, where there’s a lack of reliable publicly available speech corpora. Hence, efforts were directed towards creating a segmental vocoder with unsupervised training and the capability of performing well using training-sets of small size. Initial obtained results had unsatisfactory performance, from the quality and bit-rate perspectives, so several enhancements were investigated and were eventually used successfully. The introduced enhancements included: using shorter length segments, using segment piecewise linearization, using variable bit rate coding, and introducing two novel distortion measures: the low-band Time Domain Phase Aware CTDPA) distortion measure [1], and the high-band Energy Mass distortion measure [2, 3]. The TDPA distortion measure introduces a novel technique to jointly encode magnitude and phase information in a single codebook, avoiding the complexity of the iterative methods typically needed to create separate magnitude and phase codebooks. Also, since the TDPA is time-domain based, it showed a superior ability to maintain phase coherence and provide smooth time evolution of the synthesized speech signal, an attribute often linked to improved speech quality. On the other hand, the introduced Energy Mass distortion measure provided a means to exploit a physical property of the speech spectrum where the adjacent frequency bins were related to each other while assessing distortion. The introduced TDPA and Energy Mass |