Towards Optimal TTS Corpora

AbstractUnit selection text-to-speech systems currently produce very natural synthesized phrases by concatenating speech segments from a large database. Recently, increasing demand for designing high quality voices with less data has created need for further optimization of the textual corpus recorded by the speaker. This corpus is traditionally the result of a condensation process: sentences are selected from a reference corpus, using an optimization algorithm (generally greedy) guided by the coverage rate of classic units (diphones, triphones, words…). Such an approach is, however, strongly constrained by the finite content of the reference corpus, providing limited language possibilities. To gain flexibility in the optimization process, in this paper, we introduce a new corpus building procedure based on sentence construction rather than sentence selection. Sentences are generated using Finite State Transducers, assisted by a human operator and guided by a new frequency-weighted coverage criterion based on Vocalic Sandwiches. This semi-automatic process requires time-consuming human intervention but seems to give access to much denser corpora, with a density increase of 30 to 40% for a given coverage rate.