Leaderboard · Model
Chatterbox
A provisional entrant — listeners are still collecting blind rounds on it, so its humanness is an early estimate. Humanness 100 means listeners can no longer tell it from the person.
Standings as of August 31, 2026 · refreshed every 5 minutes
Hear it
“Thanks for calling — I can get that rescheduled for Thursday morning. Does that work for you?”
Background
Chatterbox is Resemble AI's open-source TTS (MIT license, 0.5B parameters) — the first open-weights model on this board. Resemble's own blind study reported listeners preferring it over ElevenLabs 65% of the time; the arena exists to verify claims like that independently. Clips are self-hosted (no gateway path), so latency shows a dash.
On the board
Live results — the numbers move as votes land. Latency is measured at the Speko gateway (warm first-audio, n=30); price is the provider's published rate. Method details on the methodology page; raw standings as open data (CC BY 4.0).