Blind test: 75 listeners hit 48.8% accuracy telling ElevenLabs v4 audiobook narration from human

antiochIst · reddit · 2026-11-12

A rigorously controlled blind test found listeners essentially cannot distinguish ElevenLabs v4 synthetic narration from human readers.

Key results

75 listeners judged 496 audiobook clips in a forced choice; overall accuracy was 48.8% (95% CI 44.0–53.8), d' = -0.04 — chance level.
Clip length didn't matter: 15s, 30s and 60s clips all scored 49% (60s minus 15s difference: 0.0 points).
Confidence didn't help: "very confident" answers were 48.7% accurate; of 58 high-confidence errors, 36 were AI clips called human.
The voice most often judged human was a real narrator (81%), followed by ElevenLabs' "Eric" (80%); three real narrators were recognized less than a third of the time.

Methodological highlight

AI side: six stock ElevenLabs v4 voices at default settings. Human side: ten LibriVox volunteers.
Before listening began, the author trained a small classifier on acoustic features: spectral tilt alone separated the sets with AUC 0.815 — a recording-chain difference, not a speech difference. Human clips were equalized until AUC dropped to 0.51, removing the "microphones" confound.

Limitations: interim data (41% of planned 1,200 judgments), one TTS system at default settings, one genre (calm nonfiction), 180 AI vs 30 human clips, volunteer rather than professional narrators. Not a claim that synthetic speech is undetectable in general.

Why it matters: most "humans can tell" claims come from unblinded listening or mismatched recording chains. On matched material, for this system and genre, the answer is no. Disclosure: the author runs the listener panel (Woodstitch), funded by an audiobook producer.

Original post →

Blind test: 75 listeners hit 48.8% accuracy telling ElevenLabs v4 audiobook narration from human

More from Multimodal