Acoustic and perceptual analysis of discontinuities in two TTS concatenation systems

Background Discontinuities It is fair to say that L&H’s (now Scansoft’s) RealSpeak and AT&T’s NextGen are two of the most natural sounding unit selection systems. The transitions between connected units sometimes contain discontinuities, thus creating one of the greatest problems concerning the output in these kinds of systems. The discontinuities are often perceived as ‘jumps’, i.e. a disturbance. The analyses in this paper investigate the acoustic properties of the ‘jumps’, if they are perceived as disturbing and in that case how disturbing. The results show that the selection criteria do not include enough information on single acoustic parameters, such as formants. Since listeners perceive discontinuities in formants, especially F2, as disturbing, one of the conclusions is that the next step in developing these systems must be to include more information on these parameters separately (especially formants 2 and 3) to improve the selection process. Of course other things like increasing database size and better structuring of data etc. can also improve the selection process as well as better grapheme to phoneme conversion, but those aspects are not dealt with here.