TTS and audio evaluation by native speakers, in hours.
Measure what a spectrogram can't: does this voice sound natural, friendly, right for the language? Rapidata collects MOS-style ratings and pairwise preferences on your audio from native speakers of the target language.
Naturalness, MOS and preference, the way speech teams report them.
Use the protocol your speech team already reports, or run several side by side on the same clips.
- Naturalness / MOSRate each clip on a 1–5 scale for naturalness or overall quality and average into a mean opinion score.
- Pairwise preferenceWhich of two voices sounds more natural, more friendly or more like the reference speaker.
- PronunciationNative speakers flag mispronounced words, wrong stress and unnatural prosody.
- Audio RLHFFeed the same preferences back as reward data for speech, music and sound generation models.
Judged by people who speak the language.
A Japanese voice should be judged by Japanese speakers. With Rapidata you can filter raters by spoken language and country, so each locale of a multilingual TTS model is rated by its own native audience rather than by a generic English-speaking panel.
Audio evaluations already on Hugging Face.
Both datasets were collected with the same API you would use, and every individual response is public.
Rapidata/text-2-audio-human-preference-benchmark
About 32,000 human responses collected in under an hour, asking which voice sounds more natural and which sounds more friendly.
DatasetGlobal audio studyRapidata/kiki-bouba-audio-20k
20,000 responses to spoken words from a worldwide crowd, with country, language, age and gender for every respondent.
DatasetSame crowd, same API, different shapes.
Audio & TTS evaluation FAQ
Can't find the answer you're looking for?Contact our support team at info@rapidata.ai!