Continuous mixture densities and linear discriminant analysis for improved context-dependent acoustic models

Linear discriminant analysis (LDA) experiments reported previously (ICASSP-92 vol.1, p.13-16), are extended to context-dependent models and speaker-independent large vocabulary continuous speech recognition. Two variants of using mixture densities are compared: state-specific modeling and the monophone-tying approach where densities are shared across the states relevant to the same phoneme. Results are presented on the DARPA Resource Management (RM) task for both speaker-dependent (SD) and speaker-independent (SI) parts. Using triphone models based on LDA and continuous mixture densities, significant improvements have been observed and the following word error rates have been achieved: for the SD part, 7.8% without grammar and 1.5% with word pair; and for the SI part, 17.2% and 4.6%, respectively. These scores are averaged over 1200 SD or SI evaluation sentences and are among the best published so far on the RM database.<>