A bottom-up approach for handling unseen triphones in large vocabulary continuous speech recognition

Presents an extension of bottom-up state-tying towards improved handling of unseen triphones. As opposed to the usual backing-off to diphones and monophones, the current method aims at finding a triphone model that has proven to exhibit some similarity with the unseen triphone. It is based on a probabilistic mapping of unseen contexts to clusters of triphone states observed in the training data. This algorithm has been applied to dictation tasks for three languages with vocabulary sizes ranging from 20k to 64k. The results compare favorably with those obtained using standard back-off rules. This technique also offers an alternative to top-down decision-tree procedures which are frequently used, especially for their generalization capabilities.