Supervised approach to the construction of context-sensitive acoustic prototypes

The performance of a large vocabulary speech recognition system is critically tied to the quality of the acoustic prototypes that are established in the relevant feature space(s). This is especially true when only a limited amount of training data is available to extract information about pronunciation variability. To better account for co-articulation effects, we describe a supervised strategy for the construction of context-sensitive acoustic prototypes. The idea is to incorporate contextual supervision to relate the allophonic models to their acoustic manifestations. This makes for a better utilization of the available training data, while at the same time allowing for a short design time turn around. The performance of this method is illustrated on an isolated utterance speech recognition task with a vocabulary of 20,000 words.