Abstract: In this Ph.D. dissertation the influence of packet losses on speech recognition is analyzed and different solutions to prevent, reduce and conceal their effects are developed. The performance of remote speech recognition will be subject to the robustness of the speech coding scheme used. Conventional speech codecs achieve to reduce the bit-rate by making use of predictive techniques that exploit temporal speech correlations. Thus, to decode a frame, a correct decoding of the previous ones is required. However, this inter-frame dependency reduces considerably the robustness against packet losses because it originates an error propagation in addition to the associated information loss. Furthermore, speech decoders integrate their own packet loss concealment algorithms, which are based on perceptual considerations that are unsuitable for speech recognition. In order to combat these degradations, we propose a set of mechanisms that can be divided into sender-driven and receiver-based techniques.
Index Terms: Network speech recognition, robust speech recognition, packet loss concealment.