Research And Construction Of Large Vocabulary Continuous Speech Recognition System

Large vocabulary continuous speech recognition (LVCSR) is one of the mostimportant subjects of spoken language processing, which involves many knowledgesources and techniques such as acoustic model, language model and decodingalgorithm. This paper will discuss our experience and fruits in the search of speechrecognition, and present a new decoding algorithm in LVCSR.In this paper, the construction methods of many popular language models suchas N-gram model and Latent Semantic Analysis (LSA) model are studied and severalsmoothing methods are discussed and applied successfully in speech recognition.In this paper, some large vocabulary search algorithms are studied and atree-form token passing algorithm is proposed. For the two decoding strategies ofone-pass and multi-pass, the paper achieves two different forms of tree search, usingdifferent tree and network structures to organize the partial paths and record thehistory information, which not only saves the memory but also simplifies theprocedure of decoding.In this paper, the search space control strategies of LVCSR are studied. Manydifferent pruning tips and look-ahead techniques are used to delete some unpromisinghypotheses to improve the performance of practical speech recognition system.The paper also studies how to organize and integrate language models intospeech processing to improve the recognition accuracy. For complex language model,such as Trigram and LSA, it is very expensive to use them directly in one-pass search.So we introduce a new approach named Partial-Path-Adjusting (PPA) for high-levellanguage models, which, at the time of improving the recognition performance, didnot introduce any apparent efficiency reduction.Based on these researches, we design a new speech recognition decoder on anopening platform named Gallina to achieve an efficiently large vocabulary searchprocessing. The experimental results of Gallina are satisfactory.