Best TTS model right now: the leaderboard versus Sume's router

The independent leaderboard puts Eleven v4 first for quality, and Sume's TTS router does not serve Eleven. It serves Cartesia Sonic models under the ids sonic-3.6, sonic-3.5, sonic-3, sonic-latest and sonic-preview. If a ranked first place is a hard requirement, Sume is the wrong place to generate speech today. If you want a managed TTS job that sits next to transcription, music and mixing, the router is what you get.

What the sources say

Artificial Analysis lists Eleven v4 as the quality leader with a Quality Elo of 1320, and Polly Standard as the throughput leader. ElevenLabs' changelog entry for 28 September describes Eleven v4 and v4 Turbo with support for more than 90 languages and a median latency of about 100 ms for Turbo.

How to decide

• Rank by your own script. A leaderboard averages many voices and prompts; your product reads a specific kind of text.
• Listen to Sonic output through the same endpoint you will ship. A generated sample takes one job and costs cents for a short line.
• Count the extras you need: speech-to-text, a music bed, a mixdown. Doing these in one account removes glue code.
• Check language coverage. Set language explicitly for non-English scripts on Sume.

A fair test on Sume

Call GET /v1/tts-router/models to see the current catalog, then submit the same three paragraphs to sonic-3.6 and to sonic-3.5. Compare by ear against your reference. The job echoes the model id you requested, so file names should include it. See pinning a model id.

Limits

Elo is a ranking from listener votes at the time of the read, not a guarantee for your voice or language. The leaderboard does not list Sume's job features, and this post does not claim a Sonic ranking, because none was read from the source. For cost, see the cheapest TTS API per 1,000 characters.