Abstract: In this paper we present a predictive vector quantizer for distributed speech recognition that makes use of a delayed decision coding scheme, performing the optimal codeword searching by means of the M-algorithm. In single-path predictive vector quantization coders, each frame is coded with the closest codeword to the prediction error. However, prediction errors and quantization errors of future frames will be influenced by previous quantizations, in such a way that choosing an instantaneous coding with the best codeword for each frame do not offer the optimal codeword sequence. The M-algorithm presents the advantage of obtaining a global minimization of the quantization error by maintaining the M-best quantization hypotheses for each frame, in a multipath coding approach outperforming the single-path predictive vector quantizer. In this work, the chosen cost function is the Euclidean distance between the sequence of prediction errors and the sequence of quantized values. The method has been tested for coding MFCC coefficients in Distributed Speech Recognition systems, making use of a non-linear predictive vector quantization on a large vocabulary task. Experimental results show that using this global optimization, lower bit rates can be achieved than using the single-path coding non-linear predictive vector quantizer without degradation in terms of WER.
Index Terms: distributed speech recognition, predictive vector quantizer, delayed decision coding, M-algorithm.