This paper presents a speech recognition model composed of a parallel phoneme recognition module based on continuous hidden Markov model (HMM) and a parallel sentence recognition module based on hierarchical knowledge base model. The phoneme recogniti...
This paper presents a speech recognition model composed of a parallel phoneme recognition module based on continuous hidden Markov model (HMM) and a parallel sentence recognition module based on hierarchical knowledge base model. The phoneme recognition module distributes thousands of HMMs to multiprocessors, computes output probabilities in parallel, and improves the Viterbi search. The parallel sentence recognition module generates recognition results using a prediction-activation algorithm, applied to phonetic code sequence from the phoneme recognition module and knowledge base. The algorithm is implemented on a Multi-Transputer system and a Parsytec CC, which are distributed-memory MIMD multiprocessors. In the result of experiments, we obtained improved speedup in the phoneme recognition module and improved recognition rate in the sentence recognition module. Furthermore, the experimental results show the feasibility of the real-time parallel speech recognition system based on continuous HMM.