Excitation analysis and modelling for high quality speech synthesis

Describes work which is aimed at improving the naturalness of copy synthesised female speech produced using a software implementation of the JSRU parallel formant synthesiser. The motivation behind this work is drawn from a number of recent publications which show that excitation specific effects relating to the shape of the excitation spectrum contribute significantly to the naturalness of synthetic female speech. An attempt is made to model a number of known excitation effects by dynamically changing the voice mix and mark-space ratio control signals using information obtained from a laryngographic signal. To accommodate such excitation effects, modifications are made to the bandwidth and frequency values stored in the synthesisers parameter control file. Copy synthetic speech produced using a number of different parameter combinations are compared through subjective testing and results presented