Incorporating multiple-HMM acoustic modeling in a modular large vocabulary speech recognition system in telephone environment

The use of multiple acoustic models has reported great improvements when facing speaker independent difficult tasks. In this paper, we are applying this strategy to a flexible, large vocabulary, speaker-independent, isolated-word hypothesis generation system in a telephone environment with vocabularies up to 10000 words. The new problem addressed here is how to efficiently integrate the multiple model scheme in the system, as due to its bottom-up approach (phonetic string generation followed by a lexical access process), multiple possibilities arise (apart from the alternatives in the training stage), and its not clear what combination would achieve the best results. In the paper, full details on every alternative are shown, along with results showing actual improvements in the system.