TTS leaderboard: 33 Elo points rank 5 to 12

The gap between rank 5 and rank 12 on Versely's September 2026 text-to-speech leaderboard is 33 Elo points: Gemini 3.1 Flash TTS sits at 1212 and ElevenLabs at 1179. A gap that small is a reason to audition voices on your own script rather than pick by rank.

What the leaderboard lists

The Versely write-up gives these positions. The 33-point figure is simple subtraction of the two Elo values.

Reading a 33-point spread

Versely also notes that Gemini 3.1 Flash TTS shipped on April 15, 2026 with 30 voices and multi-speaker dialogue. Those are feature facts, not quality facts. An Elo number summarizes pairwise listener votes across many prompts. It does not tell you how a model reads your product names, your language mix or your pacing.

Treat the table as a shortlist. Anything within a few dozen points of the top is worth one blind test with 3 to 5 of your real lines.

Testing a voice through Sume

Sume's hosted MCP exposes tts_create as a paid generation tool, next to stt_create and music_create. Paid calls need an idempotency_key, and dry_run=true previews admission and cost without submitting a job. Ask the server for the live contract with tools_schema and name: "tts_create" before you build on it, because the docs say not to assume parity with the HTTP API.

Which voices or engines a session can pick is listed by that schema, not by this post. Run the same three lines through each option you can select, play them back to back without labels, and keep a note of which one your listeners prefer.

A simple audition routine

• Pick lines that contain numbers, brand names and one long sentence.
• Generate each line once per voice and keep the job ids and settings next to the audio.
• Re-run the winner on a second day to check that results are stable.
• Only then compare price per character or per minute.