The first method that leads to very good results, but not in real time, was a software simulation of the linear prediction algorithm. The hardware component was only a D/A converter and the audio equipment.
We tried the formantic synthesis method to obtain speech in real-time, using the dedicated processor PCF8200 produced by PHILIPS company. The synthesizer module was an external module connected to the computer by parallel interface.
The speech quality obtained was not good enough because that chip allowed only quantified values of the parameters (formants frequencies and bandwidths). Our algorithms for computing the formants for the Romanian language diphones obtained some values which had to be adjusted for feeding the formantic synthesizer chip. This approximation leads to a loss of intelligibility in the resulted voice.
The high quality of speech obtained by the linear prediction algorithm was a reason for us to implement it on a digital signal processor, TMS320C25. This chip from Texas Instruments family can work in parallel with the computer's processor for obtaining real-time synthesis of the vocal signals.
9. Experimental results and conclusions
Our analysis-synthesis algorithm
allows the vocal signals to be decomposed into their basic components,
and to be recomposed from these components, the small distortion
being due to the mathematical imperfection of vocal tract model
used in the formant or in the LPC synthesizer and to the data
storing compression.
The system allows the modification of some characteristics
of vocal signal, in order to obtain prosodic effects (including
effects like: whispers sounds, singing sounds, hoarse sounds,
etc.).
The naturalness and the quality
of generated speech are quite acceptable, further improvements
being necessary. A morphemic analysis for grapheme to phoneme
conversion has to be tried, more rules for stress, intonation
and pitch have to be implemented.
Synthesis by rule, especially
for the Romanian language, is an open research subject.
166
References
