Multilingual TTS voice evaluation data from native listeners
Synthetic and cloned voices rated by people who grew up speaking the language: Hindi, Tamil, Marathi, Bengali, Odia and Urdu listeners, Mexican and Peruvian Spanish speakers, Arabic speakers, and speakers of Indigenous languages such as Quechua. Every clip gets a naturalness score, an accent verdict and its mispronounced words, so you know where a voice stops sounding native.
Specification
- Languages
- Hindi, Tamil, Marathi, Bengali, Odia and Urdu; Mexican and Peruvian Spanish; Arabic dialects; Quechua and other low-resource languages
- Listeners
- Native speakers from the regions where each language is spoken, balanced across gender and age groups
- Tests
- Naturalness (MOS), accent and dialect fit, intelligibility, pronunciation errors, A/B preference, and speaker similarity for cloned voices
- Material
- Synthetic and cloned clips, scored alongside native reference recordings
- Annotation
- A score per clip and listener, mispronounced words written in the language’s own script, short comments
- Quality check
- Hidden reference clips and repeated items catch careless listeners
Example of the data
What you see, and the file you get.
| Hindi · voice A | naturalness 4.3 / 5 | accent: native ✓ |
|---|---|---|
| Tamil · voice A | naturalness 3.2 / 5 | “ழ” sound wrong |
| Spanish (MX) · A vs B | B preferred | sounds Mexican |
| audio_url | language | voice | listener | naturalness_1_5 | accent_heard | mispronounced |
|---|---|---|---|---|---|---|
| …/audio/tts_0101.wav | hi | voice_a | L208 | 4 | native | |
| …/audio/tts_0101.wav | hi | voice_a | L215 | 5 | native | |
| …/audio/tts_0214.wav | ta | voice_a | L340 | 3 | non-native | வாழைப்பழம் |
| …/audio/tts_0388.wav | es-MX | voice_b | L512 | 4 | mexican | |
| …/audio/tts_0412.wav | qu | voice_c | L607 | 2 | non-native | allillanchu |
Sample rows. Also as JSON or Excel; audio as WAV or MP3.
Contact us to get the dataset
Tell us what you need. We walk you through the data, its formats and how it is checked on a short call.
What it is used for
Choosing a voice per market
Pick the voice native listeners trust in each language before you launch.
Fixing pronunciation
Get the words, names and sounds each voice gets wrong, language by language.
Voice cloning QA
Check that a cloned voice still sounds like its speaker, and still sounds native.
Tracking each release
Score every new model on the same test so progress is measured, not guessed.
Questions
Who are the listeners?
Native speakers of each language, from the regions where it is spoken: a Tamil voice is judged by Tamil speakers, a Mexican Spanish voice by Mexicans.
Which languages are covered?
Indian languages including Hindi, Tamil, Marathi, Bengali and Odia, plus Urdu, Spanish, Arabic and Indigenous languages such as Quechua. Ask us about the languages you need.
Can I see samples first?
Yes. Request samples with the form on this page and we will send them with the full specification.
Tell us what your model needs.
Request samples and the full specification. We reply within a couple of business days.
Request samples