Better TTS using LPC

Our MSR people are excited about the LPC-based TTS systems like the one developed by Keiichi Tokuda at Nagoya Institute of Technology.

The basic idea is to apply HMM to predict the spectral trajectory a sound makes, and do it at run-time based on a voice you've trained. The big advantage is smaller size compared to the concatenative systems like the new one in Vista, but it can be applied to make concatenative TTS higher quality as well.

I wish I could have met Tokuda when I was at last year's Eurospeech.