next up previous
Next: The Voice Up: Arabic in my Hand: Previous: BABYLON

The System

Cepstral's Thetasynthesizer is a unit selection synthesizer specifically developed to be optimal on a low resource device. As the BABYLON system supports two synthesizers, two recognizers and two translation engines at the same time on a device with a comparatively slow processor and only about 40M total available to our system, a small, fast engine is important.

The engine itself is designed to work on processors without floating point support; all scores and measures are done in fixed point. The core architecture of Thetais based on CMU Flite [2], but the unit selection algorithm has been optimized for space and speed.

The second important aspect of the system is that of database compression. Unit selection synthesizers require large databases to provide appropriate variation. Thus we ensure that the units within the database are the most useful ones for synthesis, and we take advantage of the fact that the voice is from a single speaker and use speech compression techniques to reduce the data representation significantly, but still ensure we can unpack efficiently at run time.



Alan W Black 2003-10-27