Rapid-deployment text-to-speech in the DIPLOMAT system
Kevin Lenzo, Christopher J. Hogan, Jeffrey Allen · 1998
The DIPLOMAT project at Carnegie Mellon University instantiates a program of rapid-deployment speech-to-speech machine translation; we have developed techniques for quickly producing text-to-speech (TTS) systems for new target languages to support this work. While the resulting systems are not immediately of comparable quality to commercial systems on unrestricted tasks in well-developed languages, they are more than adequate for limited-domain scenarios and rapid prototyping -- they generalize to unseen data with some degradation, while quality in-domain can be quite good. Voices and engines for synthesizing new target languages may be developed in a period as short as two weeks after text corpus collection. We have successfully used these techniques to build a TTS module for English, Croatian, Spanish, Haitian Creole and Korean. 1. The DIPLOMAT Project The DIPLOMAT project is an experiment in rapid-deployment, wearable, bi-directional speech translation systems. An implementation of the complete synthesis system should be available at an acceptable level of quality within a few weeks after initial recording corpus design, with continual, graceful improvement to a good level of quality over a period of months. For a small domain, the concatenative synthesis method used here can be quite adequate; however, the current synthesis solution, called phonebox, is a rapid prototyping tool rather than a substitute for a full-fledged speech synthesis system in the target language. The synthesizer receives text-only input through a socket or pipe, and can easily be replaced by an arbitrary synthesis engine and a small wrapper. We must often trade off the quality synthesis in the short term in favor of development time, memory limitations, and runtime efficiency, in order to provide basic communication technologies in contexts where communication barriers are significant, and to do so on commonly available hardware. So far, DIPLOMAT has worked with English, Croatian, Haitian Creole, Spanish and Korean.