From within-word model search to across-word model search in large vocabulary continuous speech recognition

In this paper we report on the application of across-word context dependent acoustic phoneme models in a single-pass large vocabulary continuous speech recognizer.Although across-word models are used by many groups today, only an outline of the recognizers is usually given in the publications. Implementation details are often missing.We present both a formal derivation of across-word model search and a detailed description of our implementation. The across-word model system is compared with a conventional within-word model system regarding word error rate and computational effort. Compared to the baseline within-word system a straightforward implementation of across-word model search results in a substantial increase of the computational effort. Therefore, several optimization steps are studied that result in a more efficient organization of the search space and a more efficient pruning. The effects of these optimizations are analysed in a detailed profiling. In combination they accelerate the straightforward implementation of across-word model search by nearly a factor of three.In addition we discuss the construction of word graphs during across-word model search. Starting from a word graph method based on within-word model search, we derive a formal specification of across-word word graphs. We show that the resulting word graphs are a good representation of the active search space.