Voice Arena's Turing TTS: Best AI Voice Fools Listeners 10.5%
Voice Arena pits real speakers against TTS on the same line and asks listeners to pick the human.
All ten systems stay detectable; Gemini 3.8 Flash TTS leads.
"Indistinguishable from a human" is standard copy at every speech-synthesis launch. Voice Arena's new Turing TTS leaderboard checks the claim against actual humans.
The protocol is a two-alternative forced choice. A real speaker and a TTS system read the same line; the listener picks the one they think is human. The share of votes where the AI is picked is its Fool Rate. If listeners truly couldn't tell, it would sit near the 50% expected from guessing.
The first US English results are humbling. Across ten systems, 100 lines and 2,000 counted votes, the pooled Fool Rate is 6.1%. Google DeepMind's Gemini 3.8 Flash TTS leads at 10.5%, followed by Gemini 3.8 Flash-Lite TTS at 9.0%, Cartesia Sonic-3.6 at 7.5%, and ElevenLabs Eleven v4 Turbo and Eleven v4 at 6.0% each; Fish Audio's Drama 3 trails at 3.0%. Every system's upper confidence bound is below 25%, so all land in the "Detectable" tier.
Caveats apply. These are early results with 200 votes per system, the top three's rank ranges all include first place, and a different set of reference lines could reshuffle the order. The headline holds anyway: in a head-to-head with a real voice, today's TTS is still easy to spot.
What the board really changes is the question. Preference arenas ask which voice sounds better; Turing TTS asks which sounds real. Those correlate but aren't the same—polish can be the giveaway, while breaths, hesitations and imperfections are the hardest things to fake. For teams building support agents, dubbing or companion apps, Fool Rate is closer to user perception than Elo. For fraud prevention, it is the number to watch: when it nears 50%, the alarm should go off.