A large-vocabulary real-time continuous-speech recognition system

A system architecture has been developed to implement real-time large-vocabulary continuous-speech recognition using HMM (hidden Markov model) algorithms and bigram language models. It is shown that the largest bottleneck in such a system is located in the memory access. The architecture exploits a variety of techniques, such as partitioning and replication, to cope with this memory bottleneck. The required throughput is achieved with the aid of extensive pipelining (up to thirteen levels deep) and concurrency. The architecture allows extension to larger vocabularies by the addition of more parallel units. Pin count considerations have resulted in the definition of five custom integrated circuits which are currently being tested. Using the proposed approach, the authors are currently designing and debugging a real-time 3000-word continuous-speech recognition system that uses bigram language models