Influence of Variable-Length Speech Units on Quality of Synthetic Speech Signal
Jindřich Matoušek, Daniel Tihelka, Josef Psutka · 2002
This paper examines variable-length speech units in the context of a single unit instance text-to-speech system. Experiments described here were intended as a preparation for our future multiple unit instance system. Subphoneme, phoneme and word units were studied to find out the influence of variable-length units on the synthetic speech in a fully working TTS system. Statistical approach was applied to build variable-length speech unit databases automatically. Longer units were shown to be superior because the synthetic speech contained less spectral discontinuities in comparison with shorter units. Our attention was also paid to a level of clustering speech context in the process of automatic speech unit modeling. Model-level clustering was shown to outperform state-level clustering.