What Audio Lab is
Audio Lab is a collection of audio tools that run in your browser tab instead of on a server somewhere else. It covers editing tasks like trimming, joining, and converting; effects like EQ, reverb, and bass boosting; voice and AI tools like stem splitting, transcription, and voice cloning; and creation tools like a beat maker, a granular synthesizer, and a BPM detector — 36 core tools in total, plus 2 experimental lab tools for more extreme sound design. Every tool is organized around a task rather than a technical category, so the goal is to get you from "I have an audio file and I need to do X to it" to a finished export as directly as possible.
There is no account, no watermark, and no payment anywhere on the site. You open a file, or record one, adjust the controls for the tool you're using, and download the result. Settings and presets you save stay in your own browser rather than on a server tied to an account, because there isn't one.
The principles behind it
The first principle is that processing happens on your device. Audio you open or record is decoded and processed using your browser's own capabilities — the Web Audio API for most tools, WebAssembly for format conversion, and WebGPU or WebAssembly for the handful of tools that use machine-learning models. Nothing about your audio is sent to a server for that processing to happen, which is a property of how the tools are built, not a policy choice that could quietly change.
The second principle is honesty about limits. Separating a mix into stems leaves some bleed between them. Monophonic pitch detection struggles on chords or multiple simultaneous voices. Converting a lossy file to a lossless format doesn't recover quality that's already gone. AI tools need to download a model before they can run, and that download, along with the AI processing itself, can be slow or run out of memory on an older phone or laptop. Each tool page tries to say this plainly rather than promising a result the underlying technique can't actually deliver.
How it's built
Most real-time effects — EQ, reverb, bass boosting, spatial audio — run on the Web Audio API's native filter and processing nodes. Format conversion, video audio extraction, and some metadata handling use FFmpeg compiled to WebAssembly, which is what lets Audio Lab accept a wider range of formats and containers than the browser can decode natively.
The AI-based tools run through ONNX Runtime Web and Transformers.js, which let machine-learning models execute inside the browser using WebGPU where it's available, falling back to WebAssembly otherwise. Those models are open-weight and openly licensed: stem separation and AI-mode vocal removal use Demucs v4 (MIT license), transcription uses Whisper (MIT license), text-to-speech uses Kokoro-82M (Apache-2.0 license), and voice cloning uses Chatterbox (MIT license). Model files are hosted on Hugging Face and jsDelivr, downloaded once per device, and cached by your browser afterward.
What we test
The parts of Audio Lab that are pure math get unit tests against known-correct behavior: loudness measurement should report the integrated LUFS value the calculation defines, true-peak detection should catch inter-sample peaks the way oversampled true-peak metering is supposed to, and trimming should land on the exact sample you selected rather than drifting by a few milliseconds. Testing signal-processing code this way catches regressions in the underlying maths before they show up as an EQ curve or a loudness reading that's quietly wrong.
That kind of testing has an honest boundary: it verifies that a filter or a measurement does what its formula says it should, not that every AI model output or every audio aesthetic will match your taste. This site doesn't publish a roadmap, so what you see in each tool's controls is what's actually implemented today, not a preview of something planned.