styletts2-community/common-voice-filtered

Common Voice Filtered

A filtered subset of the Common Voice dataset. Currently, this dataset only includes a small subset of English speech.

We only include speech ranked above 3.75 (75%) on the MOS metric, as calculated by the UTMOS system. Approximately 7% of audio qualified for inclusion in this filtered dataset.

This data is not final. Processing the whole Common Voice dataset would require a significant amount of compute, this is just a small sample/MVP of the project.

The code is available on GitHub.

Transcriptions

The transcriptions can be found in the original Common Voice dataset here.

Uses

Because this dataset is limited to higher-quality audio clips, this dataset is especially suitable for text-to-speech systems, however not as suitable for ASR (speech recognition) systems because it excludes lower-quality audio.

Length

Unfortunately, most phrases in this dataset are limited to 5 seconds.