Memory-intensive recognition for word articulation training

Deaf children may be helped to improve their articulation by using discrete utterance recognition techniques to match the words they pronounce with many template words pronounced by hearing children. This requires a multiple-speaker system, to accommodate the wide intra- and inter-speaker variations in "correct" pronunciations. With many different templates, time-warping techniques such as dynamic programming can lead to excessive computation time, and require very fast processing units for interactive applications. The system described here uses a large memory to hold the templates, and a very simple algorithm for recognition. By using recognition during training, and retaining a new template only if it is not recognized when compared with existing ones, it aims at capturing intra-speaker and inter-speaker pronunciation variations rather than attempting to model them mathematically. This approach can therefore be broadly characterized by two aspects: 1. Trade-off in favor of memory size versus processing speed. 2. Use of the recognition algorithm from the beginning of the system training phase.