Large Vocabulary Recognition

We give the results of experiments in which we apply our method for speaker dependent transcription of speech to the speaker independent case. We have retrained our existing system on the TIMIT data base withholding utterances of one male and one female speaker for testing purposes. Our acoustic/phonetic model was trained by fitting Gaussian densities to 12th order delta-cepstra, energy and differential energy of observations of each of 51 phonetic units based on the labeled TIMIT data. Gamma densities were also fitted to the measured segmental durations. The state transition matrix was estimated in a novel way.