Research on narrowband communications

This document reports on work toward a very low rate vocoder. We model speech as a Markov Chain of spectral templates for the unsupervised learning approach to very low rate vocoding. This quarter we compared several clustering techniques. We determined that a hierarchical approach to clustering is economical with minimal loss in performance. Furthermore, we found that a small number of spectral templates (from 128 to 256) is sufficient for vocoding with good intelligibility. Also, in the phoneme recognition approach, we automated the training of the diphone network on labelled speech.