VoiceArena just gave AI voices a test they actually fail. Their new Turing TTS… · @harden_369741 · TwKit

@harden_369741

VoiceArena just gave AI voices a test they actually fail. Their new Turing TTS asks listeners to pick the human out of a human vs model pair, and the best model on the board, Gemini 3.8 Flash TTS, fooled people only about 10% of the time. I think that number is the honest one. We keep ranking voices by which clip sounds nicer in a 30 second blind vote, and that leaderboard says the field is nearly solved. A fool rate says the opposite. The top four models are statistically tied on preference, yet none of them passes as human nine times out of ten. My read: naturalness leaderboards flatter the models, fool rates humble them, and for anything where a voice has to hold a real conversation, I would trust the humbling number. #AI #TTS #VoiceAI

曝光 21 · 评论 0 · 点赞 0 · 书签 0