This repository collects time alignments for popular TTS datasets, removing leading and trailing pauses. Additionally, a sentence is split when a pause is > 0.3 seconds.
We use the https://montreal-forced-aligner.readthedocs.io/en/latest/ (MFA) and extract, for each file, the following information:
path|text|speaker|start|end
where start and end are given in seconds. For example:
VCTK-Corpus/wav48/p295/p295_138.wav|I haven't had a lesson in four years.|VCTK_p295|0.6898185941043085|2.569818594104308
Removing leading and trailing silences is standard practice in text-to-speech synthesis, and these files should help.
Currently, we supply alignments for
- VCTK
- LibriTTS
- Blizzard
If you want to contribute, please submit a PR.
If you found my alignments useful, built something cool or have any questions, please drop me a line or submit an issue!