Supertonic: The Lightning-Fast, Fully On-Device TTS That Actually Works in 2025

Core Question: What exactly is Supertonic, and why is it running 100–167× faster than real-time on a laptop or phone — completely offline?

Supertonic is a 66-million-parameter text-to-speech (TTS) model released by Supertone in 2025. Built for extreme on-device performance and powered by ONNX Runtime, it runs 100% locally on everything from smartphones to browsers — no cloud, no API keys, no privacy trade-offs. With just 2 inference steps it already sounds production-ready, and on Apple M4 Pro it hits an insane 167× real-time speed.

Why Supertonic Changes Everything: Real-World Use Cases

Core Question: When would you actually need a fully local TTS this fast?

Here are four situations where existing solutions simply fail:

• Building an offline reading or audiobook app that must work in airplane mode while keeping user data private.
• Creating in-game NPCs that speak player-written dialogue (including numbers, prices, dates) with near-zero latency.
• Developing browser extensions for screen readers that cannot phone home.
• Upgrading smart-speaker firmware with a TTS model under 100 MB that still sounds natural.

Supertonic solves all of them at once: tiny footprint, instant generation, native handling of numbers/dates/currencies/abbreviations, and zero text preprocessing required.

Personal Takeaway
After years of testing open-source TTS (Piper, Coqui, Kokoro, etc.), most either crawl on laptops or butcher “$99.99” and “2025-11-21”. The first time I heard Supertonic read a raw sentence on my M4 Pro in 0.006× real-time, I literally paused for five seconds. That “it just works” moment is incredibly rare in local AI.

Supported Platforms: Literally Everywhere You Code

Core Question: Which platforms can run Supertonic natively today?

The official repo ships ready-to-run examples for 10 ecosystems:

No matter your stack, you’re covered.

Performance Numbers That Speak for Themselves

Core Question: How fast is Supertonic compared to ElevenLabs, OpenAI, and other open models?

Official 2-step inference benchmarks (2025 hardware):

Characters per Second (higher = better)

Real-time Factor (lower = better)

On an RTX 4090, long text hits RTF 0.001 — that’s 1,000× real-time. You could turn live chat messages into spoken commentary faster than humans can talk.

Get Your First Voice in Under 30 Seconds

Core Question: How do I hear Supertonic right now?

Step 1: Clone code + model

git clone https://github.com/supertone-inc/supertonic.git
cd supertonic

# You need Git LFS for the model files
git lfs install
git clone https://huggingface.co/Supertone/supertonic assets

Step 2: Quickest Python demo

pip install onnxruntime soundfile tqdm

python py/inference.py \
  --text "Hello world! Today is November 21, 2025. I paid $99.99 for coffee." \
  --voice preset/ko-KR \
  --output hello_supertonic.wav

Play hello_supertonic.wav — numbers, dates, and currency pronounced perfectly, generated in a blink.

Step 3: Instant browser demo (WebGPU)

Just visit the official Space:
https://huggingface.co/spaces/Supertone/supertonic

Or run the web example locally — works in Chrome/Edge 121+ out of the box.

Photo by Pexels

Advanced Tuning: Squeeze Every Last Millisecond

Key command-line flags you’ll actually use:

--steps         # 2 = fastest, 5 = slightly more natural
--batch_size    # 4–16 for very long texts
--temperature   # 0.6–0.8 sweet spot
--cfg_weight    # default 1.0 works great

Real-world sweet spots I found:

• M4 Pro WebGPU + --steps 2 → stable RTF 0.006
• Android (Java example) + --steps 3 → best speed/quality trade-off
• RTX 4090 + --batch_size 16 → 12k+ chars/sec

Deep Dive: Four Killer Applications

1. Offline Audiobook / Read-Aloud Apps
   A 5,000-character article synthesizes in under 4 seconds on M4 Pro — entirely offline.

2. Real-Time Game Dialogue
   Player types “Give me 3.14 apple pies for €42.00” → NPC responds instantly in natural voice.

3. Privacy-First Browser Screen Readers
   Works on any webpage without ever sending text to a server.

4. Always-On Smart Speakers
   Local weather reports, reminders, or jokes even when Wi-Fi is down.

My Two Cents
What blew me away wasn’t just the raw speed — it was the zero-preprocessing text handling. No more writing regex hell to make “$99.99” sound right. That alone makes Supertonic production-ready for most developers.

Licensing & Commercial Use

• Code: MIT → do whatever you want
• Model weights: OpenRAIL-M → commercial OK, just don’t use for illegal/harmful stuff

One-Page Cheat Sheet

Quick-Start Checklist

1. Install Git LFS
2. Clone repo + assets
3. pip install dependencies (Python) or open the web folder
4. Run a demo → hear magic
5. Port to your target platform

FAQ

Q: Does Supertonic support Chinese voices yet?
Pre-trained presets are currently Korean/English-focused, but the architecture is multilingual and fine-tuning scripts are coming.

Q: Will it run smoothly on phones?
Yes — official Java (Android) and Swift (iOS) examples run 3–5 step inference fluently on 2023+ flagship devices.

Q: How does quality compare to ElevenLabs?
2-step is close to ElevenLabs Flash; 5-step reaches much of their Prime tier — entirely on-device.

Q: Which browsers support the WebGPU demo?
Chrome 121+, Edge 121+, Safari on macOS 15+.

Q: Is commercial use allowed?
Absolutely — MIT code + OpenRAIL-M weights.

Q: Why is a 66M model this fast?
Aggressive distillation + ONNX quantization + a novel 2-step diffusion design.

Q: Any voice cloning?
Current release is zero-shot multi-speaker only. Cloning tools may appear later.

Q: How can I contribute or train my own voice?
Watch the official GitHub — training code is on the roadmap.

Supertonic isn’t just another TTS model. It’s the first open on-device solution that finally feels like the future we were promised: instant, private, and actually usable. Go try the demo right now — you’ll hear the difference in under ten seconds.