Thunder Clarity
v0.1In developmentEvery recording carries the room it was made in. Thunder Clarity pulls the voice, its own reverb tail, and the ambience apart into three layers — then hands you a fader for each, in real time.
The three faders
Turn the room down. Keep the voice.
Noise gates and de-reverbs work by removing — and you hear the removal, as pumping, as gating, as a voice that breathes in and out of a hole. Thunder Clarity does not remove anything. It separates first, then lets you rebalance what came apart.
The separated dry voice. Push it up to bring dialogue forward — this is the core of de-noising, done by lifting rather than by cutting.
The voice's own reverb tail, on its own fader. Pull it down to de-reverb a roomy take; push it up when the natural space is the point.
Everything else in the room — traffic, hum, crowd, air. Pull it down to de-noise, or push it up to keep the scene and isolate the room instead.
A diagram, not a recording — no audio is being processed on this page. It shows the one thing that matters about separation: pull Ambience down and the floor between words drops while the voice stays exactly where it was. In the shipping Phase-1 build that is measured at −11 dB.
Each fader is a balance, not a gate. You are remixing separated stems, so there is no pumping and no gating artifact when a layer comes down — and unity is bit-exactly transparent. Built for streamers, post-production, podcasts and musicians.
The engine
What we stand on, and what we built.
The separation core is a real-time neural speech enhancer. We did not reinvent the model from zero — we stand on openly-licensed research and put our work where it actually differentiates.
DeepFilterNet
Real-time full-band speech enhancement — de-noise and de-reverb — with pretrained models. MIT / Apache-2.0, a Rust core behind a C API.
RNNoise
Lightweight RNN noise suppression from Xiph, BSD-licensed. It is the low-latency path, for when you are monitoring live rather than mixing.
Everything you touch
The stem-rebalancing DSP, the reverb and ambience split, the fault-tolerant real-time host integration, and the plugin itself.
Both cores are permissively licensed and commercially usable with attribution, which we carry in the product. Naming them is not a disclaimer — it is how you can check what is actually ours.
Where it stands
Phase one works. We will not pretend the rest does.
The streaming spectral separator is in and running: pull Ambience down and it de-noises, with −11 dB measured on the noise floor in the gaps between words while the voice stays where it was. The deep neural core and the dedicated reverb stem are still being built.
- Working now
- Real streaming-STFT separation. Voice and Ambience faders. Bit-exact transparency at unity.
- Next
- The neural core, for deeper suppression than a classic separator can reach.
- After that
- The dedicated Voice Reverb stem, and a selectable low-latency path for live monitoring.
- Foundation
- Openly-licensed research engines rather than a model reinvented from zero — with the same input-integrity guard and fail-safe reversion proven in Sibilance.
- Formats
- VST3 · Audio Unit · CLAP · Standalone. JUCE 8 · C++20.