In this paper we will present some experiments to improve the naturalness of synthetic female speech. A pole/zero pair has been introduced in the glove synthesiser to better match analysis data. In this paper we have focused on the consonants /h/, /l/, /m/ and /n/. An inverse filtering method, which includes automatic formant tracking and manual spectral and time domain matching, generates IF source parameter, formant and pole/zero data. Analysis data have been interpreted and resynthesised. Text-to-speech rules have been formulated according to the hand-edited resynthesis. These rules have improved the overall quality of the synthetic female speech according to an informal listening test.

Keywords: hi-fi female speech synthesis, acoustic description of consonants