| Summary: | Voice communication systems are moving at a steady pace from circuit-switched to packet-switched networks. Converged networks, in which data traffic and multimedia traffic share the same infrastructures, have yet to face many challenges. Unlike data traffic, voice traffic is very sensitive to issues such as delay, jitter and packet losses. The Quality of Service (QoS) of IP networks has greatly improved over the recent years. Nevertheless, it is not uncommon to experience loss in QoS on some networks caused by lack of resources or software/hardware failures. The speech codec (COder DECoder) is an essential component in a digital voice communication system. It converts the speech signal into a compact representation in order to efficiently use the available bandwidth. To compress the signal, speech codecs usually exploit the high redundancy present in a speech signal using techniques known as predictive coding. The most widely used speech coding model is the Code Excited Linear Prediction (CELP) model upon which are based most of the modern standardized speech codecs. The use of predictive codebooks, the Adaptive CodeBook (ACB) in particular, allows a tremendous reduction in bitrate. Unfortunately, the extensive use of prediction renders CELP codecs more vulnerable to packet losses than non-predictive codecs. A possible solution is to avoid using prediction at the cost of an increased bitrate. In this thesis, we propose several techniques to reduce the vulnerability of CELP codecs to packet losses caused by the use of prediction. Solutions that aim at limiting the contribution of the ACB and thus that reduce the sensitivity of the codec to packet losses are presented and discussed. Techniques based on sending side information, in order to help the decoder recover after a packet loss, are also proposed. The solutions are constrained to be low delay and low bitrate solutions, and are required to maintain interoperability if used in a standardized codec. Objective and subjective test results that demonstrate the performances of the proposed solutions are presented. We show that the robustness of CELP codecs can significantly be improved without sacrificing their compression performances. In spite of packet loss issues, we show that it is still beneficial to use predictive coding to efficiently exploit the available bandwidth.
|
|---|
