Decagon Chord's 45.7% human test: what the number measures

Decagon reports that 45.7% of listeners believed the real human voice was the AI, which it describes as approaching the 50-50 threshold. That is a listening test on Chord in customer conversations, run by the vendor. It does not show that a voice will sound human in a 15 second product ad, and the vendor's own Voice 3 post gives a different headline (about 90% of listeners could not tell Chord from a human), so quote the figure with its page.

Numbers are from the two Decagon pages read on 2026-10-13; I did not read Decagon's test protocol beyond those posts. The Voice 3 page puts the blind-test figure at about 90% across three voice samples, so the two numbers are not the same measurement.

Why not reuse the number for an ad voice?

Decagon frames the test around customer conversations, where it says pacing shifts with the length and meaning of a sentence and speech carries pauses and fillers. An ad read is scripted, short and heavily edited. Your audience also hears it over music with captions on screen, so a different test fits.

How do I run my own blind test on Sume?

Render the same script with two router models and the same language, then compare. Both jobs use the same request shape; only model changes. Sonic and Eleven rows carry different voice sets, so pick one voice per family from GET /v1/tts-router/voices. Keep the take lengths identical so loudness and pace do not decide the vote.

run() {
  curl -s -X POST https://api.sume.com/v1/tts-router/generate \
    -H "Authorization: Bearer $SUME_API_KEY" \
    -H "Content-Type: application/json" \
    -H "Idempotency-Key: blind-$1-001" \
    -d "{\"model\":\"$1\",\"transcript\":\"Your 15 second ad line.\",\"voice\":{\"id\":\"$2\"}}"
}
run sonic-3.6 "$SONIC_VOICE_ID"
run eleven-v4-turbo "$ELEVEN_VOICE_ID"