Synthesizer Voice Quality of New Languages Calibrated with Mean Mel Cepstral Distortion
John Kominek, Tanja Schultz, Alan W. Black · 2008
When developing synthesizers for new languages one must select a phoneset, record phonetically balanced sentences, build up a pronunciation lexicon, and evaluate the results. An objective measure of voice quality can be very useful, provided it is calibrated across multiple speakers, languages, and databases. As a substitute for full listening tests, this paper adopts mean mel-cepstral distortion as a measure of spectral accuracy, and proposes systematic variation of a known English corpus as a method of calibration. We find that doubling the database size reduces MCD by 0.12, while reverting to a grapheme-based voice increases it by 0.27. This offers a frame of reference for estimating voice quality, which is applied to a test suite of 8 non-English languages. 1.