Third International Conference on Spoken Language Processing (ICSLP 94)
We report on some recent improvements to a continuous density hidden Markov model (CDHMM) based speech recognition system which is being developed for a variety of telecommunication applications. In particular, we are concerned with computing the emission probabilities of an HMM state using a combination of multi-layer perceptrons (MLPs) as probability estimators and mixture-Gaussian densities. Using MLPs as state observation estimators has been shown to improve accuracy on a speaker verification task . In this paper, we describe a connected digit recognition system which incorporates both MLPs and mixture, Gaussian densities. The results are reported on the standard Texas Instruments (TI) connected digit database  which was digitally filtered to the telephone bandwidth (300 Hz-3.2 kHz) and downsampled to 8 kHz. A hybrid MLP/HMM system led to 15% improvement in performance. The final string error rate is 1.7% for unknown length strings.
Bibliographic reference. Lubensky, David M. / Asadi, Ayman O. / Naik, Jayant M. (1994): "Connected digit recognition using connectionist probability estimators and mixture-Gaussian densities", In ICSLP-1994, 295-298.