A Large-Vocabulary Continuous Speech Recognition Algorithm and Its Application to a Multi-Modal Telephone Directory Assistance System

A LARGE-VOCABULARY CONTINUOUS SPEECH RECOGNITION ALGORITHM AND ITS APPLICATION TO A MULTI-MODAL TELEPHONE DIRECTORY ASSISTANCE SYSTEM Yasuhiro Minami, Kiyohiro Shikano, Osamu Yoshioka, Satoshi Takahashi, Tomokazu Yamada and Sadaoki Furui NTT Human Interface Laboratories 3-9-11 Midori-cho, Musashino-shi, Tokyo, 180 Japan ABSTRACT This paper describes an accurate and efficient algorithm for very-large-vocabulary continuous speech recognition based on an HMM-LR algorithm. The HMM-LR algorithm uses a generalized LR parser as a language model and hidden Markov models (HMMs) as phoneme models. To reduce the search space without pruning the correct candidate, we use forward and backward trellis likelihoods, an adjusting win- dow for choosing only the probable part of the trellis for each predicted phoneme, and an algorithm for merging candi- dates that have the same allophonic phoneme sequences and the same context-free grammar states. Candidates are also merged at the meaning level. This algorithm is applied to a telephone directory assistance system that recognizes spon- taneous speech containing the names and addresses of more than 70,000 subscribers (vocabulary size is about 80,000). The experimental results show that the system performs well in spite of the large perplexity. This algorithm was also ap- plied to a multi-modal telephone directory assistance system, and the system was evaluated from the human-interface point of view. To cope with the problem of background noise, an HMM composition technique which combines a noise- source HMM and a clean phoneme HMM into a noise-added phoneme HMM was investigated and incorporated into the system. 1. INTRODUCTION One of the main problems with very-large-vocabulary con- tinuous speech recognition is how to accurately and effi- ciently reduce the search space without pruning the cor- rect candidate. Our speech recognition system is based on the HMM-LR algorithm [1] which utilizes a generalized LR parser [2] as a language model and hidden Markov mod- els (HMMs) as phoneme models. Applying this algorithm to large-vocabulary continuous speech requires: (1) accurate scoring for phoneme sequences, (2) reduction of trellis calcu- lation, and (3) efficient pruning of phoneme sequence candi- dates. For the first requirement, several speech recognition algo- rithms that calculate the backward trelhs likelihood from the end of the utterance, as well as the forward trellis likehhood, have been proposed [3][41 . We also use forward and back- ward trellis likelihoods for accurate scoring. For the second requirement, we use an adjusting window, which chooses only the probable part of the trellis according to the predicted phoneme. For the third requirement, we use an algorithm for merging candidates which have the same allophonle phoneme sequences and the same context-free grammar states [5]. In addition, candidates are also merged at the meaning level [6]. Speech HMMs are sensitive to incoming noise and this of- ten results in a large decrease in the recognition. One solu- tion is to train HMMs on noisy speech to obtain the corre- sponding optimum HMMs. For large-vocabulary continuous speech recognition, however, the computation load of this so- lution becomes too high, because all the HMMs need to be re- trained each …

Content truncated to 500 words