Next: Search Algorithm
Up: Special Features
Previous: Efficient Models
The task of decoding is to find the maximum likelihood word sequence given the models and acoustic evidence. The time synchronous Viterbi algorithm provides an efficient means of performing this task for small vocabularies (e.g., less than 1000 words) and short span language models (e.g., bigrams). However, with larger vocabularies and longer span language models a simple exhaustive search is not possible and the issue of efficient decoding becomes critical to the performance of the system.