Evaluation and Analysis of Auditory Model Front Ends for Robust Speech Recognition Program Summary

This project was motivated by the need for improved speech recognition in noise, and by expectation that auditory model front ends could make recognition more robust to noise, microphone variation, and speaking style. The project began in FY91 with the initial goal of implementing, evaluating, and comparing four promising auditory front ends: (1) the mean-rate and synchrony outputs of S. Seneff's auditory model; (2) the ensemble interval histogram (EIH) model developed by O. Ghitza; (3) the IMELDA model due to M. Hunt; and (4) an auditory model developed by J. Cohen. The initial effort has included implementation and integration of the first three models with a Lincoln HMM recognizer, and testing with speech in white noise and in a background of speech babble. with white noise and speech babble background. As compared with the mel-cepstrum technique, the auditory models have similar performance at high SNR and slightly better performance at low SNR. In order to select the best features from each front end, dimensionality reduction techniques, including Linear Discriminant Analysis (LDA) and Principal Components Analysis (PCA) have been implemented. LDA has been applied to obtain a transformation matrix for IMELDA, with the transformation matrix developed using 100,000 labeled speech frames (1,000 sentences) from the TIMIT corpus. Testing of IMELDA is about to begin on the TI-105 corpus.