Whole word phonetic displays for speech articulation training.

摘要

The main objective of this dissertation is to investigate and develop speech recognition technologies for speech training for people with hearing impairments. During the course of this work, a computer aided speech training system for articulation speech training was also designed and implemented. The speech training system places emphasis on displays to improve children's pronunciation of isolated Consonant-Vowel-Consonant (CVC) words, with displays at both the phonetic level and whole word level. This dissertation presents two hybrid methods for combining Hidden Markov Models (HMMs) and Neural Networks (NNs) for speech recognition. The first method uses NN outputs as posterior probability estimators for HMMs. The second method uses NNs to transform the original speech features to normalized features with reduced correlation. Based on experimental testing, both of the hybrid methods give higher accuracy than standard HMM methods. The second method, using the NN to create normalized features, outperforms the first method in terms of accuracy. Several graphical displays were developed to provide real time visual feedback to users, to help them to improve and correct their pronunciations.