An AI music mixer and browser drum machine: twelve tracks, a 32-step sequencer and fully synthesized sounds, running entirely in your browser.
Make beats in your browser.
A modern groovebox - no signup, no uploads, no plugins.
Twelve tracks, one browser tab
Pick a demo pattern, press play, then make it yours: edit the drum steps, mix levels and panning, push the tempo.
A step sequencer on the Web Audio API
Everything you hear is computed in your browser at the moment you press play.
All lanes are editable. Drum steps cycle through normal, accent and soft velocity, and the hi-hat lane cycles closed and open. Bass, chords and lead open a scrollable piano roll: click to set a note or a chord (pick a root and a chord type), drag right to stretch its length. A pattern library lets you clone, rename and build variations of a groove, and a song lane arranges patterns into full songs by drag & drop - saved in your browser, exportable as self-contained files.
Timing does not come from JavaScript timers. A lookahead scheduler plans every event 120 milliseconds ahead on the audio clock, so playback stays sample-accurate even when the browser throttles the page.
Every sound is synthesized from first principles - a kick is a sine wave with a falling pitch envelope, a snare is a filtered noise burst over a triangle body. There are no third-party samples anywhere in this application, which also means there are no third-party licences.
The mixer needs no account, records no audio and never uploads your music: a network trace of a full session shows only GET requests for the page's own assets and model files - no audio or musical data is sent anywhere.
An on-device pattern engine
The interesting part of this project is machine learning that runs where your music is: in the browser.
A small variational autoencoder (VAE) - about 276,000 parameters, roughly 540 KB in fp16 - generates the drum patterns. It was trained on symbolic drum data, not audio: it predicts which drum hits which step and how hard. Generate samples a new pattern near a style's anchor in the model's latent space; Variation encodes the pattern you are currently playing and nudges it by the amount you choose, so small settings change a few steps and large ones approach a fresh roll of the dice. Inference runs in this browser tab via TensorFlow.js; no audio or musical data is sent to any server.
Honest limits, measured on the test split: the model reproduces held-out patterns with an onset F1 around 0.70, it under-generates open hi-hats and toms because they are rare in its training genres, and quantising human playing to a sixteenth-note grid costs about 7 percent of the original onsets. The full engineering story - what the AI tooling built well, where it failed, and what stayed human work - is being written up as a technical article on our research channel.
The pattern model was trained on the Groove MIDI Dataset (Google LLC), licensed under CC BY 4.0: Jon Gillick, Adam Roberts, Jesse Engel, Douglas Eck, David Bamman, "Learning to Groove with Inverse Sequence Transformations", ICML 2019.
Browse our research channelWhy a mixer lives at /mixer.html
This address has a history that predates us - and that history is exactly why this project exists.
The mybytes.com domain had a prior life before today's myBytes GmbH acquired it. Historical sources show that /mixer.html once hosted a browser-based music application operated as part of a Microsoft education campaign about intellectual property. Today's myBytes GmbH has no organisational or commercial connection to that former offering - we are neither its legal successor nor a continuation of it.
When we found the dormant URL, we asked a more interesting question than "what was here": how would you build a browser music application today, and how much of that work can modern AI-assisted engineering actually carry? This page is the living answer, built from scratch with today's web platform - Web Audio, workers and on-device machine learning - and documented as an engineering case study.
Curious about the original? The Internet Archive holds snapshots of this URL from 2008 to 2015, independent of us.