Research
Blinded listening studies on speech AI, run with paid human raters. Each report states its sample size, its confidence intervals and what it does not show. We publish results that go against us.
Which ElevenLabs voice is best for audiobooks?
32 Voice Library voices, chosen to span gender, age, accent, recording tier and popularity, read the same 30-second nonfiction passage. 125 listeners judge them blind, in pairs. Listen to the voices yourself, then see where each finished.
95% CI 55–76
studio tier no better
95% CI 16–35
125 listeners · 32 voices · 1 passage · 1,122 judgments · popularity predicts nothing · louder clip wins 54%Read the report →
Research report 6 November 2026 · Hawkes TWhich TTS model is best for audiobooks?
ElevenLabs v4, Gemini 3.8 Flash TTS, Fish Audio Drama 3, Cartesia Sonic 3 and Qwen 3.1 TTS Flash read the same sixteen passages from classic novels. 78 listeners judge them blind, in pairs. Listen to the readings yourself, then see who read them.
95% CI 49–63
order not established
95% CI 34–48
78 listeners · 5 systems · 16 passages · 481 judgments, preregistered sample reachedRead the report →
Interim research report 5 November 2026 · Hawkes TWhich AI model writes the best piano music?
Eight large language models compose short piano loops for the same twenty prompts, all played on one piano. 62 listeners judge them blind. Listen to the loops yourself, then see who wrote them.
95% CI 57–75
95% CI 51–70
95% CI 19–35
62 listeners · 8 models · 20 prompts · 514 of 1,638 judgmentsRead the report →
Interim research report 3 November 2026, updated 6 November · Hawkes TCan listeners distinguish AI narration from human narration?
A blinded listening study of ElevenLabs v4 and human-read audiobook excerpts of 15, 30 and 60 seconds.
95% CI 44.0–53.8
no length effect detected
confidence did not track accuracy
75 listeners · 6 AI voices · 10 narrators · 210 clips analysedRead the report →