Reducing the natural click effect within database for high quality corpus-based speech synthesis
Monique V. Nicodem, Rui Seara, Fernando Santana Pacheco · 2006
This paper presents a technique for detecting and smoothing involuntary clicks generated by the human vocal tract, which degrade the quality of text-to-speech systems. This approach is useful for high quality corpusbased concatenative speech synthesis. The proposed click detection technique is based on bandpass ltering the low energy subbands of a speech signal, autoregressive modeling, prediction error, and a thresholding approach. By using a windowing technique in the smoothing phase one reduces considerably the undesired click effects. Preliminary experimental results verify the applicability of the proposed approach.