Menu

Dataset: 1,122 blind pairwise human judgments on 32 ElevenLabs TTS voices (paired preference, CC BY 4.0)

Reading 0:00
15s threshold

We needed a defensible way to pick narrator voices from the ElevenLabs Voice Library (5,000+ English narration voices, no quality signal), so we ran a paid blind test.

Setup: 32 voices stratified across gender, age, accent, tier (professional vs instant clone) and usage; the same 30-second first-person nonfiction passage; identical eleven_v4 settings for every voice; 128 pairs in a connected circulant design; 125 paid US listeners, 1,122 forced-choice judgments; no listener heard any voice more than twice. Bradley-Terry ranking with listener bootstrap, plus a logistic property model on the voice tags.

Results:

Marie - Audiobook Narrator and Measured finished first (66% of pairings won, 95% CI 55–76). Alder - Warm, Raspy and Comforting last (24%). The top eight overlap and can't be ordered.

The "professional" badge told listeners nothing. Studio-recorded PVC voices won 46% of 649 head-to-head pairs against instant clones.

Popularity predicts nothing.…

Continue reading — create a free account

Join HashtagPLUS to read full articles, follow hashtags, vote, and join the conversation.