Quantization of cepstral parameters for speech recognition over theWorld Wide Web

摘要

We examine alternative architectures for a client-server model ofnspeech-enabled applications over the World Wide Web (WWW). We compare anserver-only processing model where the client encodes and transmits thenspeech signal to the server, to a model where the recognition front endnruns locally at the client and encodes and transmits the cepstralncoefficients to the recognition server over the Internet. We follow annovel encoding paradigm, trying to maximize recognition performanceninstead of perceptual reproduction, and we find that by transmitting thencepstral coefficients we can achieve significantly higher recognitionnperformance at a fraction of the bit rate required when encoding thenspeech signal directly. We find that the required bit rate to achieventhe recognition performance of high-quality unquantized speech is justn2000 bits per second