We report on research toward a very-low-rate vocoder. This quarter we continued investigation in two areas: the phonetic vocoder and an unsupervised method for vocoding. We introduced phoneme pair probabilities to improve the accuracy of phonetic recognition. We investigated the use of the phonetic vocoder as a tool for semi-automatic labeling of speech. We also experimented with several variations of the phonetic vocoder to improve the intelligibility of the vocoded speech with a moderate increase of the bit rate. Our work in the second area concentrated on developing a novel method of spectral quantization: We quantize a sequence of spectra simultaneously. We are also evaluating a new distance metric that does not require the dynamic programming time warping.
