Voice Arena's Turing TTS test: best TTS model fools listeners about 10% of the time
Article timeline
The articles in this story. Times are ET.
Rohan PaulXVoice Arena's Turing TTS benchmark: best model fools listeners 10% of the timeAIVoice Arena launches Turing TTS, a benchmark where listeners hear a human and a model say the same sentence and pick the human. Gemini 3.8 Flash TTS, the top model, fooled listeners 10% of the time, and the top four models on its preference board are statistically tied.
SantiagoXTuring TTS benchmark finds listeners mistake Gemini 3.8 Flash TTS for humanAISantiago (@svpino) says a new benchmark measures how often listeners mistake AI speech for human speech in paired recordings. He names Gemini 3.8 Flash TTS as the best model at fooling people, while admitting he is easily fooled himself. The background post introduces Turing TTS, where the best model fooled listeners only about 10% of the time.
elvisXVoice Arena's Turing TTS tests how often listeners mistake AI speech for humanAIVoice Arena introduces Turing TTS, an evaluation measuring how often human listeners mistake AI-generated speech for real human speech. In its tests of leading TTS models against human speech, the best model fooled listeners only about 10% of the time. The post argues the community should examine whether blind pairwise naturalness evaluations are enough to measure progress in voice AI.