Trellis encoded vector quantization for robust speech recognition

A joint data (features) and channel (bias) estimation framework for robust speech recognition is described. A trellis encoded vector quantizer is used as a pre-processor to estimate the channel bias using blind maximum likelihood sequence estimation. A sequential constraint in the feature vector sequence is explored and used in two ways, namely, a) the selection of the quantized signal constellation, b) the decoding process in joint data and channel estimation. A two state trellis encoded vector quantizer is designed for signal bias removal applications. Compared with the conventional memoryless VQ based approach in signal bias removal, the preliminary experimental results indicate that incorporating sequential constraint in joint data and channel estimation for robust speech recognition is advantageous