This paper describes an algorithm/or very large vocabulary (about 80,000 words) continuous speech recognition. To improve recognition accuracy and reduce the amount of computation, we implement an accurate and efficient algorithm based on a two-stage LR parser with phoneme HMMs. In this algorithm, we adopt the forward and backward trellis likelihoods for accurate scoring. We also adopt methods of adjusting windows and merging in phoneme sequences as well as grammatical states for efficient searching. This algorithm is applied to a telephone directory assistance system containing the names of more than 70,000 subscribers. The system recognizes the continuously uttered names and addresses. The results show that the system has good performance in spite of the large perplexity.

Keywords: Continuous Speech Recognition, /IMM, Search Algorithm