Efficient distribution of feature parameters for speech recognition in network environments

Abstract

In network or ubiquitous environments, there are difficulties in performing large vocabulary speech recognition by a small device due to its limited power. Therefore, an approach, so-called distributed speech recognition (DSR), that; distributes the processing modules of automatic speech recognition into a device and a server has been attractive. Of all processing modules of DSR, quantization of feature parameters plays a main role in terms of the transmission bandwidth and the recognition performance. In this paper, we propose an efficient quantizer of feature parameters by incorporating the correlation between successive analysis frames of speech. The proposed quantizer is based on the predictive multi-stage vector quantization scheme and designed with different bit rates by trading off with the performance of speech recognition. It is shown from speech recognition experiments that the DSR system employing the proposed quantization method can reduce a bit rate by 20% with a comparable recognition performance to the ETSI DSR standard.