Benchmark
Controlled Voice Arena
Compare text-to-speech models through listener preferences using the same cloned voices.
Published leaderboard
Explore scores and evaluation settings on the publisher’s leaderboard.
Open leaderboardCoverage
Eight professionally recorded English reference voices and two native-speaker reference voices per additional language. Current methodology covers English and nine other languages.
Methodology
Listeners compare anonymous, randomized clips generated from the same text and cloned reference voice. Screened pairwise preferences produce Bradley-Terry ratings, with separate fits for languages and other subsets and 95% confidence intervals.