LDec: One Pass Time Synchronous Decoder

Authors: Tomás Pavelka, Václav Matoušek

Abstract:

The search for the most probable word sequence is in automatic speech recognition called decoding and is usually carried out by the Viterbi algorithm, an efficient search strategy based on dynamic programming. The paper discusses implementation issues and methods to further reduce computational costs when performing recognition with large vocabularies and stochastic language models. Such methods can be divided into two categories: those that eliminate unnecessary computations (such as tree structured lexicons) and those that exclude computations according to the probability that those computations will lead to the desired result (pruning).