The Effect of SNR and GCI Perturbation on Speech Synthesis with Harmonic plus Noise Model

Harmonic plus noise model (HNM) divides the speech spectrum in two bands: harmonic and noise. As most of the non-periodic components are removed in harmonic part, it may be expected that HNM synthesis is less susceptible to additive noise in input speech. As HNM analysis/synthesis is performed at each glottal closure instant (GCI), errors in estimation of GCIs affect the quality of the synthesized speech. Effects of the amount of additive broadband noise in input speech and perturbation in GCIs on the synthesized speech quality with special reference to phoneme sets in Indian languages were studied. Synthesis results show that for SNR in the 2-10 dB range, the quality of synthesized speech is superior to that of the input speech. Investigations also show that the speech quality is very sensitive to positions of the GCIs. Perturbations above 8 % severely affect quality of the output speech.