ASample ANot played
Many languages. One listening community.
A world of voices.
Human preference, language by language. Select a language to explore scores or contribute a vote.
Colored cells show evidence, not winners. Ratings are comparable only within a language and comparison group. No multilingual average.
Loading language coverage…
EVALUATION RESULTS
The leaderboard.
Compare languages
Select any languages. Leave empty to follow the Arena language.
Equal-weight language averages, not universal quality scores. ASR difficulty differs by language; WER and CER are never mixed. A metric stays undefined if any selected language lacks it. Human preference remains per-language.
Early results are useful clues, not final winners. Bands show approximate uncertainty; overlapping scores may change order. Overall uses observed tasks only when some tasks lack votes.
| Model | Human preference ⓘ | RTF ↓ ⓘ | WER ↓ | Speaker similarity ↑ | Clips | License |
|---|---|---|---|---|---|---|
| Loading verified results… | ||||||
Beta · — Undefined or unmeasured. CER replaces WER for unsegmented languages. API RTF includes network and provider processing: useful for delivery speed, not a controlled L4 compute comparison. Open weights do not necessarily mean an open-source license.
MULTILINGUAL VOICE ARENA
Two voices. You decide.
Listen to both. Pick your favorite.
TARGET TEXT
Load a pair to start listening.
Listen to the reference voice
BSample BNot played
Audio loads only when requested. Please allow a moment for the first playback.
Try three comparisons in a language you speak. Ties are welcome.
ABOUT YODALINGUA
Find the voice you would actually use.
YodaLingua compares text-to-speech models on the same reference voices and texts. Your listening preferences help people choose a model for their language.
—models
—reference voices
—generated clips
1. Listen
Pick a language you understand. Play A and B. They read the same text using the same reference voice.
2. Choose
Pick your preferred delivery, a tie, or skip. Names stay hidden until you vote. No account required.
3. Explore
Compare human preferences and automatic measurements. Share a chart or try three comparisons to help the next listener.
What do the scores mean?
Human preference: which delivery listeners choose. WER/CER: estimated word or character errors; lower is better. Speaker similarity: how closely the voice matches its reference. RTF: generation time per second of audio; lower is faster.
Natural · everyday conversationExpressive · meaning and moodPronunciation · clear, correct words
Why do some scores say “Early results”?
They use limited listening evidence. More votes can change the order. Uncertainty stays visible, and models without enough connected evidence remain unranked. A score of 1000 is a group baseline, not a quality percentage. Scores from different languages cannot be averaged.
Methodology, limitations & privacy
Each language uses 20 speakers and 60 distinct evaluation texts. Outputs are frozen; hallucinations are retained. Automatic measurements can make mistakes. Draft translated texts need independent review, and training contamination and accent coverage remain limitations. API RTF includes network/provider time and is not a controlled local-GPU speed comparison.
Early ratings require five judgments, three sessions and three speakers per model within one connected group. Regularized Bradley–Terry estimates use approximate uncertainty bands; these are not guaranteed confidence intervals. Observed tasks have equal total weight; absent tasks are not invented. Stronger evidence adds session/speaker resampling. Exact settings are documented in the repository.
No advertising trackers, microphone recording or routine CAPTCHA. Anonymous session storage, listening checks and rate limits discourage casual abuse; sessions are not verified people. The Space is public and delivered audio can be saved. No individual votes are included in chart exports.
Read the code and full methodology ↗