An efficient algorithm for combining vector quantization and stochastic modeling for speaker-independent speech recognition

An algorithm for speaker-independent isolated/ continuous speech recognition based on stochastic modeling and vector quantization (VQ) techniques is presented. In this, the VQ technique is combined and integrated into the stochastic model such that it allows a nonparametric representation within the parametric stochastic model. The word models thus obtained are represented as a sequence of high-fidelity segmental sound units called "sound-kernels." The integrity of these models is verified by resynthesizing the words. Thus, the interaction between recognition and synthesis has been one of the main themes of this research to obtain improved recognition models. The paper also presents an improved version of a VQ design algorithm to obtain robust kernel codebooks. The overall performance of the algorithm is being evaluated with the NBS-TI multidialect connected digit data base, and preliminary results of this investigation are also reported.