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.
Can 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 42.2–56.1
no length effect detected
lowest of four levels
35 listeners · 6 AI voices · 10 narrators · 210 clips analysedRead the report →
Study 29 October 2026ElevenLabs v4 vs a fine-tuned open model, on audiobook narration
20 matched passages from 20 books, heard from both systems by every listener, sides counterbalanced.
127 decisive judgments · 6 listeners · we built the losing systemRead the study →
Collecting now
| Study | Question | Judgments | Listeners | Status |
|---|---|---|---|---|
| AI vs human narration | Continuation of the report above to its planned 1,050 judgments | 176 | 35 | Collecting |
| TTS engine preference | Do listeners prefer any of five current engines? | 202 | 37 | Collecting |
| Voice preference, audiobooks | Which voice properties predict what listeners choose? | 48 | 8 | Collecting |
| Instrumental loop preference | Which model composes the piano loop listeners prefer? | 147 | 25 | Collecting |
Counts as of 4 November 2026. Each study is reported at a sample size fixed before collection; anything published earlier is labelled interim. Null results are published the same as any other. How studies are run →