Supertonic 2

A 66M-parameter on-device text-to-speech system from Supertone, running on ONNX Runtime, licensed OpenRAIL-M (free-voice-clone-list). Five languages (English, Korean, Spanish, Portuguese, French), streaming, and no voice-cloning.

The speed claim, as published

The list reproduces the project’s own comparison table:

Stated as “up to 167× realtime” and “outperforms ElevenLabs Flash v2.5 by up to 42× in speed.”

This is a vendor table and the wiki holds it as one. It compares a local GPU against elevenlabs’ hosted API, which means it is measuring different things — one number includes no network, the other does, and neither condition is stated. The kokoro row is the part to distrust most: 117 chars/sec at RTF 1.3 would make this wiki’s efficiency leader slower than realtime, which no other source here reports. Recorded as a competing claim rather than folded into kokoro‘s page, per the record-don’t-overwrite rule. A neutral latency benchmark remains the spoke’s standing open question, and this does not answer it.

Two things that make it an outlier in the catalogue

No cloning. In a list titled free-voice-clone, 35 of 36 TTS entries claim zero-shot cloning and Supertonic 2 is the one that doesn’t — at the small, fast end, which is where open-weight-tts‘s efficiency-versus-controllability trade has always shown up. (The same list credits kokoro with cloning, contradicting this wiki’s other sources; that dispute is recorded on free-voice-clone-list.)

OpenRAIL-M. It is the only TTS entry under a licence carrying behavioural use restrictions, in a field where 27 of 36 are Apache-2.0. The list says nothing about why, and one case supports no conclusion — but it is the only place in 59 KB of voice-cloning catalogue where use restrictions appear at all (see free-voice-clone-list on that silence).

Related

kokoro · kittentts · open-weight-tts · text-to-speech · voice-cloning · elevenlabs · tts-benchmarks · free-voice-clone-list