A one pass decoder design for large vocabulary recognition

Abstract

To achieve reasonable accuracy in large vocabulary speech recognition systems, it is important to use detailed acoustic 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 necessary 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 timesynchronous 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 preliminary passes and is relatively efficient in implementation. It was included in the HTK large vocabulary speech recognition system used for the 1993 ARPA WSJ evaluation and experimental results are presented for that task.

Efficient lattice rescoring using recurrent neural network language models

“…Efficient use of language models in speech recognizers [20,19,17] requires that the context dependent states representing different histories during search can be appropriately shared among multiple paths. This applies to both conventional back-off n-gram and feedforward NNLMs.…”

The Application of Hidden Markov Models in Speech Recognition

“…To deal with this, a number of different architectural approaches have evolved. For Viterbi decoding, the search space can either be constrained by maintaining multiple hypotheses in parallel [173,191,192] or it can be expanded dynamically as the search progresses [7,69,130,132]. Alternatively, a completely different approach can be taken where the breadth-first approach of the Viterbi algorithm is replaced by a depth-first search.…”

Evaluation of a language model using a clustered model backoff

“…However, for continuous speech recognition systems using higher order language models the linguistic state cannot be determined locally and the word boundaries are uncertain. Several solutions based on creating copies of the PPT for each unique linguistic context solve this problem [8,9,10], however these approaches create redundant sub-tree computations, the number of which correspond to the number of active linguistic contexts. A computation is redundant when a sub-tree instance is dominated by another instance of that sub-tree.…”