DISTRIBUTED MANDARIN SPEECH RECOGNITION UNDER WIRELESS ENVIRONMENT

With the rapid development of wireless communications, it is highly desired for users to access the network information with spoken dialogue interface via hand-held devices at any time, from anywhere. One possible approach towards this goal is to perform speech feature extraction at the hand-held devices (the clients) and have all other recognition tasks absorbed by the server. This paper investigated problems that this scenario encounters. A “phonetically distributed” Mandarin speech database including all possible Mandarin syllables and context relationships with frequencies roughly proportional to those occurring in daily Mandarin conversation is used to train a best set of vector quantization (VQ) codebooks, such that the syllable recognition accuracy degradation due to quantization errors is minimized. We then discuss the effect of random errors and use extrapolation to compensate for it. Experimental results indicated that under our VQ scheme effect of quantization errors is slight, and at medium or low error rates we can ignore the effect of random errors, even at high error rates we can also ignore it by using extrapolation for error concealment.