A One Pass Decoder Design For Large Vocabulary Recognition

A One Pass Decoder Design For Large Vocabulary Recognition J.J. Odell, V. Valtchev, P.C. Woodland, S.J. Young Cambridge University Engineering Department Trumpington Street, Cambridge, CB2 1PZ, England ABSTRACT To achieve reasonable accuracy in large vocabulary speech recognition systems, it is important to use detailed acous- tic models together with good long span language models. For example, in the Wall Street Journal (WSJ) task both cross-word triphones and a trigram language model are neces- sary to achieve state-of-the-art performance. However, when using these models, the size of a pre-compiled recognition network can make a standard Viterbi search infeasible and hence, either multiple-pass or asynchronous stack decoding schemes are typically used. In tl:fis paper, we show that time- synchronous one-pass decoding using cross-word triphones and a trigram language model can be implemented using a dynamically built tree-structured network. This approach avoids the compromises inherent in using fast-matches or pre- liminary passes and is relatively efficient in implementation. It was included in the HTK large vocabulary speech recog- nition system used for the 1993 ARPA WSJ evaluation and experimental results are presented for that task. 1. INTRODUCTION Hidden Markov Models (HMMs) have been used suc- cessfully in a wide variety of recognition tasks ranging from small isolated word systems assisted by heavily constrained grammars to very large vocabulary uncon- strained continuous speech systems. Part of the success of HMMs is due to the existence of computationally ef- ficient algorithms for both the training of the models (the Baum-Welch algorithm) and for the decoding of un- known utterances (the Viterbi algorithm). However, as recognition tasks have become more complex, the decod- ing process has become more difficult due to the increas- ing size of network needed in a conventional Viterbi de- coder. In particular, using cross-word triphones or long span language models can lead to order of magnitude increases in the size of a static network. A variety of schemes have been proposed to reduce the computation required for recognition [2,7]. Most make use of fast-matches or preliminary passes using simpli- fied acoustic or linguistic models to constrain the search space for a final pass that uses the most detailed and accurate models available. Unfortunately, these pre- liminary matches can introduce errors that subsequent passes are unable to correct. If the first pass could use the best acoustic and language models available it would allow greater constraints to be placed on the search space without increasing the error rate. This paper describes a scheme which allows this through the use of a dy- namically built tree-structured network. This approach avoids the compromises inherent in using fast-matches or preliminary passes and is both simple and efficient to implement. The remainder of this paper is organised as follows. Sec- tion 2 discusses the main features of conventional time- synchronous Viterbi decoding and some of the ways in which it can be improved. In section 3, a one-pass de- coder that implements a beam-pruned Viterbi search through a tree-structured dynamic network is then de- scribed. Section 4 …

Content truncated to 500 words