Discriminative training of a connected digit recognizer with fixed filler models and its application to telephone network service systems

The hidden Markov models computed using discriminative training procedures have improved connected digit speech recognition accuracy significantly. The introduction of filler models to filter the extraneous speech and noise in a speech response has improved the robustness of speech recognizers. A procedure to compute digits models which share filler models with various applications running on the same system is presented in this paper. These models are used for the recognition of isolated digits and connected digits of known and unknown lengths. A process model to implement these models on a field system supporting multiple network based applications is described in this paper.