No abstract

No abstract

This paper p-t s a new timedomain algorithm for text-to-speech synthesis using diphone concatenation. The algorithm is based on the pitch-synchronous overlap-add (PSO-LA) approach, and is capable of good quatity prosodic modifications of natural speech. The algorithm can be seen as a simplif'icatio~~ of a p i o u s algorithm combining the PSOLA approach and fquencydomain uansfonnations. On the other hand, it appears as a gcnemlhtion of previous timedomain methods that perform pitch-synchronous cut-and-splice operations on the specch waveform.This algorithm is used in the CNET diphone synthesis multilingual system, actually supporting three langages: Pmnh, Italian and German. The resulting speech has been tested on Fmnch and is judged of much better quality than for an LPC-based synthesizer.

A new frequency domain method for determining the fundamental frequency of speech is presented in this paper. This method uses the information contained in the short-term phase spectrum whereas the previous methods were limited to the amplitude spectrum. The short-term spectrum is computed by DFT and it is interpreted as the output of a bank of band-pass overlapping filters. Harmonic components are detected by searching for sets of three contiguous filters having the same instantaneous frequency.

No abstract