Isolated word recognition based upon source coding techniques

Abstract

We describe an application of a recently developed speech compression technique to automatic recognition of isolated words (from a given dictionary). The scheme maps sampled speech, from a given word, into a finite codebook (of the same size as the dictionary) using linear predictive coding (LPC) all-pole models and a minimum distortion or nearest neighbor rule between all samples from the given word and the codewords from the codebook. Standard LPC techniques are used to design the codebook, but the final system requires no on-line LPC analysis.