Supertonic, Fact-Checked: The Benchmarks Are Real, the Spec Sheet Isn't Current

Supertonic just killed ElevenLabs.
A text-to-speech model that runs entirely on your device. No cloud. No API key. No per-character pricing.
2,700 GitHub stars. 100% open source. MIT licensed.
The numbers are wild:
→ 167x faster than real-time on an M4 Pro
→ Only 66M parameters
→ 1,263 chars/sec vs ElevenLabs Flash at 287
→ 1,048 chars/sec vs OpenAI TTS-1 at 55
→ Runs on a Raspberry Pi. Runs on an e-reader in airplane mode.
Reads currency, dates, phone numbers, and technical units correctly without preprocessing. ElevenLabs fails these. OpenAI fails these. Gemini fails these.
Supports 11 platforms and 5 languages. Chrome extension turns any webpage into audio in under a second.
I've watched on-device models lose to cloud APIs for years. This one doesn't lose.
The cloud TTS business just got cooked.

Now I have solid numbers to check against. Here's the fact check:

✅ Confirmed as stated

Claim	Verdict
Runs entirely on-device, no cloud, no API key	Supertonic's own benchmark documentation describes it as running via ONNX with browser support and batch processing, entirely local inference.
167x faster than real-time on M4 Pro	On M4 Pro WebGPU, long-input RTF hits about 0.006 — roughly 167× faster than real-time.
Only 66M parameters	Matches — but only for Supertonic 2, not the current flagship. The GitHub repo's own version table lists Supertonic 2 at ~66M parameters and 5 languages, while the current default (Supertonic 3) is 99M parameters and 31 languages.
1,263 chars/sec vs ElevenLabs Flash at 287	On long inputs (266 chars), Supertonic on M4 Pro CPU hits 1,263 chars/sec versus ElevenLabs Flash v2.5's 287. Exact match.
1,048 chars/sec vs OpenAI TTS-1 at 55	On mid-length inputs (152 chars), Supertonic on M4 Pro CPU hits 1,048 chars/sec versus OpenAI TTS-1's 55. Exact match — though note this is mid-length, while the 1,263 figure above is long-length; the post blends two different input-length columns without saying so.
Runs on Raspberry Pi / e-reader in airplane mode	The repo includes a Raspberry Pi demo video and an Onyx Boox Go 6 e-reader demo in airplane mode achieving an average RTF of 0.3× with zero network dependency.
Reads currency, dates, phone numbers, technical units correctly; ElevenLabs/OpenAI/Gemini fail	The README's test table shows Supertonic passing financial expressions, phone numbers, and technical units, while ElevenLabs, OpenAI, Gemini, and Microsoft all fail the same cases.
11 platforms	The repo lists ready-to-use examples across Python, Node.js, Browser, Java, C++, C#, Go, Swift, iOS, Rust, and Flutter — 11 platforms.
Chrome extension, webpage to audio under a second	A "Built with Supertonic" project (TLDRL) turns any webpage into audio in under one second on-device.

⚠️ Two claims need real correction

1. "5 languages" is stale. 5 languages was true for Supertonic 2 (paired with the 66M-parameter figure), but the current flagship, Supertonic 3, expanded language coverage from 5 to 31 languages as of its April 2026 release. The post is quoting Supertonic 2's spec sheet (66M params + 5 languages) as if it describes the current model — internally consistent with itself, but it's describing a superseded version, not what you'd get by default today.

2. "MIT licensed" is incomplete — and this is the biggest issue. Only the sample code is MIT licensed; the model weights themselves are released under the OpenRAIL-M license, which carries use-based restrictions (OpenRAIL licenses typically prohibit certain harmful uses). Calling the whole project "100% open source, MIT licensed" glosses over the fact that the actual model — the part that does the work — isn't MIT.

⚠️ Star count is badly stale

2,700 stars is far below reality: the repo currently sits at 11.2k stars, 1.2k forks, and 86 watchers. Given the project was trending at roughly 719 new stars per day at points during its growth, 2,700 was likely accurate only for a very brief early window — this number is now off by roughly 4x.

Bottom line: ~70% Real / 30% Fluff. The performance numbers (throughput, RTF, text-normalization wins) are precise and verifiable — genuinely impressive, not exaggerated. But the post cherry-picks the older, smaller model's specs (66M params, 5 languages) while implying they describe what you'd get today, understates the license (OpenRAIL-M on the actual model, not MIT), and the star count is roughly 4x out of date. The "killed ElevenLabs" framing is also just opinion dressed as verdict — a legitimate take on the benchmark, but not itself a checkable fact.