Free HD Text-to-Speech in Your Browser — No Upload, No Cloud
Quick TTS just added a third neural voice engine: Supertonic HD. It's studio-grade, 44kHz audio, and it runs entirely on your own GPU through WebGPU — nothing you paste is ever sent anywhere. Here's what it is, what it sounds like, and where it still falls short.
What Supertonic HD is
Supertonic HD is a free neural text-to-speech voice based on Supertonic 3, an open-weight model released by Supertone Inc. It runs in the browser via WebGPU and onnxruntime-web, the same "runs on your machine, not our servers" model Quick TTS already uses for Kokoro. It's the highest-fidelity engine we offer: 44kHz output, versus 24kHz for Kokoro, which is the difference between "clearly synthetic but pleasant" and audio that sits closer to a produced voiceover.
It joins Web Speech (your OS's built-in voices), Piper (small, WASM, runs anywhere), and Kokoro (WebGPU, English-only) as the fourth option in the engine dropdown — not a replacement for any of them. If you want the honest breakdown of when to reach for which, our Web Speech vs Piper vs Kokoro comparison still holds; Supertonic slots in above Kokoro on quality and below it on device compatibility.
How it runs without a server
The model downloads once — about 380MB, pulled directly from Hugging Face — and is cached by the browser afterward. From that point on, synthesis happens entirely on your GPU. No text you paste, upload, or generate audio from ever leaves your device; there's no request to log, because there's no server in the loop at all.
- ~99M parameters, roughly the same size class as Kokoro-82M, but trained to produce full-rate 44kHz audio instead of 24kHz.
- WebGPU-only. There is no WASM fallback — we measured Supertonic's WASM path at roughly realtime speed on a high-end desktop CPU, which is unusable on anything weaker, so we didn't ship it. That means a modern desktop browser: Chrome or Edge 113+ with a working GPU.
- No WebGPU, no problem. Devices without it — most phones, older laptops, Firefox and Safari today — keep Piper and Web Speech exactly as before. Nothing about the existing engines changes.
Ten voices, no cloning
Supertonic ships ten preset voices — five male (M1–M5) and five female (F1–F5) — selectable from the same voice picker as Kokoro and Piper. There's no voice cloning and no way to upload a reference sample; what you get is what's in the model. If character consistency across a long project matters more than raw fidelity, that's a real constraint worth knowing going in.
Auto language: one voice, mixed-language text
The feature that doesn't exist anywhere else in Quick TTS: an "Auto language (mixed text)" toggle. Flip it on and Supertonic reads a document that switches languages mid-paragraph — an English article quoting Japanese dialogue, a French email with an English product name — pronouncing each part correctly without you tagging anything. Every other engine we ship assumes one language per pass; this one genuinely doesn't care where the language boundary falls.
It's a narrow use case, but for anyone reading multilingual notes, bilingual study material, or code-switched social posts aloud, it's the one thing on this page that no competing free tool does.
Speed: HD quality without the wait
The tradeoff you'd expect from higher-fidelity audio is slower generation, and that isn't really true here. On a WebGPU-capable machine we measured a real-time factor of roughly 0.02–0.07 — in one test, a 13-minute audiobook-length chunk of text generated in about 13 seconds. Playback itself starts in well under a second, and long documents are chunked and pipelined the same way Kokoro's are, so a full chapter doesn't mean a full chapter's wait before you hear anything.
Which languages it covers (and the one it doesn't)
Supertonic HD is available in 15 languages on Quick TTS: English, Spanish, French, German, Portuguese, Italian, Russian, Polish, Dutch, Turkish, Arabic, Vietnamese, Hindi, Japanese, and Korean. It handles every file format and workflow the rest of the site does — PDF, DOCX, EPUB, chunked long-document playback, and WAV/MP3 download, with a ZIP export for very long texts.
It does not cover Chinese. That's worth being upfront about, since Chinese is one of our larger audiences: for zh-cn, Piper's huayan voice remains the neural option, and it's a solid one — see our Chinese TTS voices roundup if you're deciding between offline options for Mandarin specifically.
Limitations
The honest list, not the marketing one:
- 380MB is a real download. On a slow or metered connection, that's the biggest cost of trying Supertonic — bigger than Kokoro's ~80MB WASM path or Piper's ~60MB voices.
- WebGPU is required, no exceptions. No fallback path exists on this engine specifically, unlike Kokoro which degrades to Piper/Web Speech automatically when the GPU can't cope.
- Desktop-first. Mobile GPUs and mobile browser WebGPU support are inconsistent enough that we're not recommending it there yet.
- Ten fixed voices, no cloning. If you need a specific voice identity, this isn't the tool.
- The upstream project moves fast, then stops. Supertone Inc.'s reference implementation on GitHub was archived shortly after we integrated it. The model weights and inference code we ship are pinned to a specific revision and verified by us — not silently tracking an upstream that might not get further updates.
Licensing, for anyone building on this: the inference code is MIT (ported from Supertone's sample code), and the model weights are OpenRAIL-M — a separate, more restrictive license than MIT that includes use restrictions. We serve the full text at /licenses/supertonic-openrail-m.txt, and our Terms page covers acceptable use if you're generating audio for redistribution.
Try it
Open Quick TTS on a desktop with Chrome or Edge, switch the engine dropdown to Supertonic HD, and paste anything. The first ~380MB download happens once; after that it's cached and offline-capable like Kokoro. If your browser or device doesn't support WebGPU, the dropdown simply won't offer it — Piper and Web Speech keep working exactly as they did yesterday.
For the wider picture on free TTS quality tradeoffs, Best Free TTS Voices in 2026 is the place to start, and our Kokoro WebGPU benchmarks post has the methodology if you want to reproduce numbers like these on your own hardware. The comparison page covers where Quick TTS's whole free, in-browser model stacks up against paid cloud TTS.