Abstract: The most widespread techniques for speech synthesis and voice conversion are currently based on probabilistic frameworks. Particularly, Hidden Markov Models (HMMs) play a relevant role in speech synthesis, whereas Gaussian Mixture Models (GMMs) are almost standard in voice conversion. Consequently, in both cases the performance of the systems is limited by three main factors: 1) the suitability of the statistical models; 2) the over-smoothing phenomenon; 3) the accuracy of the underlying speech parameterization and reconstruction method. This paper focuses on the third issue, still open at present: translating speech frames into parameter vectors with good properties for the mentioned statistical frameworks, and reconstructing waveforms properly. The proposal presented in this paper uses the Harmonics plus Noise Model (HNM) to extract MFCC+f0 and reconstruct speech frames from them. The results of a perceptual evaluation show that the tool is valid for state-of-the-art HMM-based speech synthesis systems.
Index Terms: speech parameterization, statistical parametric speech synthesis, voice conversion, harmonics plus noise model.