Glossary · Models & serving
How does an Elo rating rank voices?
Elo rating (Voice Arena)
An Elo rating ranks voices by blind head-to-head votes. Listeners hear two clips of the same text without knowing who made them, pick the better one, and every vote nudges the two ratings apart.
On a call
A listener hears two voices read "आपकी EMI इस शुक्रवार को due है" and picks the second because the first sounded flat on "शुक्रवार". Across thousands of such votes, the leaderboard settles into a ranking.
AI agents: ₹5/min to start · ₹1.50/min all-in at volume.¹
Explained
Here is how it works, in plain words.
Arpad Elo designed his rating system for chess players. Win against a stronger player and you gain a lot of points. Beat a weaker one and you gain a little. Over many games, the ratings settle where they predict who will win.
Voice leaderboards borrow the idea. Instead of games, they collect votes. A listener hears the same sentence read by two voices, picks the one that sounds better, and that pick counts as a win. Voice Arena, like Chatbot Arena for language models, fits the ratings with a statistical model called Bradley–Terry.
Blind voting is the point. Brand names and price tags can’t sway a listener who doesn’t know which clip came from whom. What’s left is how the voice sounds to ordinary people.
Read a leaderboard with three things in mind.
Ratings are relative. A score means nothing alone. Only the gap between two voices tells you how often one beats the other.
Margins matter. A rating comes with a ± figure, its uncertainty. Few votes give a wide margin. If two voices’ margins overlap, the ranking between them could flip next week.
Categories differ. A voice can lead on conversation and trail on news reading. Check the category that matches your use, and listen to the samples in your own language.
Arenas also move. New voices join, votes keep arriving, and ratings drift. A screenshot from three months ago is history, so quote the date you pulled it.
None of this replaces your own test. Play candidate voices your real script, on a real phone line, to people who sound like your customers.
In Voice Infra
This is what Voice Infra does about it.
Leaderboards like this rank voices from blind votes, and the margin matters as much as the rank. We don't quote one for Mira TTS today. The fair test is your own lines, played blind to people who speak the language every day.
- Comparison
- Looking for a Sarvam alternative?
Questions
People often ask us these questions.
What does the ± after an Elo rating mean?
It's the margin of uncertainty, usually a confidence interval. A voice with few votes has a wide margin, so two voices whose margins overlap may not really differ.
Still deciding? Ask us.
Fifteen minutes with the team that builds it. Bring your script, or your hardest call.
Book 15 minutes