An implementation and evaluation of two diphone-based synthesizers for Turkish

This paper presents two diphone based Turkish text-to-speech systems; the first system is realized inside the MBROLA project, a freely available multilingual speech synthesizer and the second system is based on shape invariant harmonic modeling. Both synthesizers use the same parametric representations of two diphone databases (male, female) obtained by processing speech data with a pitch asynchronous, fixed frame length MBE analyzer. To obtain a pitch synchronous representation from the original asynchronous representation for the harmonic synthesizer, harmonic phases are submitted to a phase shifting algorithm, which also estimates maximum harmonic frequencies for each frame based on the evolution of harmonics’ phases. Text to phonetics transcription is achieved by one-to-one mapping of letters to phonemes. Informal listening tests have been performed for quality assessment.