Improving quality of MBROLA synthesis for non-uniform units synthesis

The paper describes a new version of the MBROLA (multiband resynthesis overlap add) algorithms (Dutoit, T. and Leich, H., Speech Commun., vol.13, p.435-40,1993) for the synthesis of non-uniform units (NUU). This new version is called TP-MBROLA, standing for true-period MBROLA. The database preprocessing of MBROLA has been modified such that short-time speech frames are not systematically resynthesized at a constant pitch and constant phase envelope. This operation highly reduces the coding-decoding effect on signal quality. For spectral smoothing, only the smoothing frames are resynthesized at constant pitch and phase envelope and MBROLA smoothing is applied. Furthermore, these operations are performed on the fly, which brings some computational load to the synthesis (though it is restricted to smoothing frames). The new version of MBROLA is tested on nonuniform units synthesis by synthesizing speech with units provided by LIMSI's unit selection system (Prudon, R. and d'Alessandro, C., Proc. 4th ISCA Speech Synthesis Workshop, p.137-42, 2001). Formal listener tests have shown that TP-MBROLA synthesis quality is preferred compared to MBROLA and raw concatenation synthesis.