Speko's Arena
372 blind votes & counting

Which voice sounds most human?

round 1/10

Voice 1
Voice 2

Play both voices to unlock voting · space to play
3 of these 10 rounds hide a real human — spot them; call the rest all-AI.

372
blind votes
11
models ranked
35%
tricked when a real human was hiding
measured live on 147 human rounds

The leaderboard

The line every voice is chasing

Humanness runs 0–100. The human sits at 100; an AI reaches it only when listeners can no longer tell them apart.

View the full leaderboard
Humanness · English live
·
Humanreference
100
1
MiniMax Speechspeech-2.6-hd
95
2
xAI Grok Text-to-Speechtts
92
3
ElevenLabseleven_v3
89
4
Hume Octaveoctave-2
88
5
OpenAI Speechgpt-4o-mini-tts
86
7
Alibaba Qwen-TTSqwen3-tts-instruct-flash
76

How it works

Ten seconds a round. Real statistics underneath.

01

Listen, blind

Two anonymous voices read the same line — no names, no logos, only audio.

02

Pick the more human voice

Even when both are AI, one usually sounds more human. Pick the one closer to a real person.

03

See the ranking

Each vote feeds a Bradley-Terry model with 95% intervals, anchored to a real human.

Get the 2026 humanness report

Enter your email to get the report when it ships, plus a heads-up when the ranking moves.

  • The full 2026 humanness report (PDF) when it ships
  • An email when the ranking shifts — new voices, protocol changes
  • Per-voice failure maps from listener tags, once collected

Every voice here runs on Speko

Speko is the voice-AI gateway — one API for ElevenLabs, OpenAI, Cartesia and every provider above, auto-picked by benchmark for latency, cost and naturalness.

Build with Speko

The arena

Think you can tell human from machine?

It takes ten seconds a round. Your ears help rank every voice in the arena.

Play a round