midicat.org stands on a great deal of work that other people made and then generously shared. This page lists all of it, and the people behind it. Thank you, every one of you.
- HT-Demucs (Hybrid Transformer Demucs), 6-instrument model
- Simon Rouard, Francisco Massa and Alexandre Défossez, at Meta AI Research. Demucs was created by Alexandre Défossez.
The model that takes a finished song apart into vocals, drums, bass, guitar, piano and everything else. Described in the paper “Hybrid Transformers for Music Source Separation” (2022).
MIT licence. github.com/facebookresearch/demucs - Browser-ready copies of the model
- published on Hugging Face by the accounts kramp (htdemucs-6s-webgpu-onnx) and StemSplitio (htdemucs-6s-onnx)
The same model converted to the ONNX format so that a web browser can run it. midicat downloads one of these the first time you convert a song.
Shared by their authors on Hugging Face. huggingface.co
- Basic Pitch
- Rachel M. Bittner, Juan José Bosch, David Rubinstein, Gabriel Meseguer-Brocal and Sebastian Ewert, at Spotify's Audio Intelligence Lab
The model that listens to one instrument and says which notes are sounding. Described in the paper “A Lightweight Instrument-Agnostic Model for Polyphonic Note Transcription and Multipitch Estimation” (ICASSP 2022).
Apache 2.0 licence. github.com/spotify/basic-pitch
- ONNX Runtime Web
- Microsoft and the ONNX Runtime contributors
The engine that runs both models inside the browser, on your graphics card (WebGPU) or your processor (WebAssembly).
MIT licence. onnxruntime.ai - jsDelivr
- the jsDelivr team and its sponsors
The free content delivery network that serves ONNX Runtime Web to your browser.
www.jsdelivr.com - Hugging Face
- Hugging Face, Inc.
Hosts the separation model files that midicat downloads.
huggingface.co
- SpessaSynth
- Spessasus
The SoundFont synthesizer and MIDI sequencer that plays the finished MIDI live on the page (version 3.24).
MIT licence. github.com/spessasus/SpessaSynth - stbvorbis.js
- Hajime Hoshi, built on stb_vorbis by Sean Barrett
Unpacks the compressed sounds inside the SoundFonts; included with SpessaSynth.
Apache 2.0 licence. - GeneralUser GS
- S. Christian Collins
A General MIDI and Roland GS compatible SoundFont. One of the sound sets on the player.
GeneralUser GS licence v2.0. schristiancollins.com/generaluser.php - FluidR3 Mono
- Frank Wen (the original Fluid R3 SoundFont), Michael Cowgill (the mono version), Ethan Winer (temple blocks), Michael Schorsch (drumline percussion)
The classic open General MIDI SoundFont, as shipped with MuseScore 2.
MIT licence. - MuseScore General
- S. Christian Collins, building on FluidR3 Mono by Frank Wen and Michael Cowgill, with Ethan Winer and Michael Schorsch; published by MuseScore
MuseScore's own General MIDI SoundFont (the file “MS Basic”).
MIT licence. musescore.org - TimGM6mb
- Tim Brechbill
A complete General MIDI SoundFont in under 6 MB; the one MuseScore 1 came with.
GNU General Public License, version 2.
The parts that are midicat's own: finding the tempo, the swing, the time signature and bar 1; reading the drum pattern; tidying the notes, snapping them to the grid and turning slides into pitch bends; choosing the instruments; writing the MIDI file; the built-in and chip sounds; the level meters in the banner; and this page you are looking at. Design and direction by Chai Webs.
None of this would run in a browser tab without WebGPU, WebAssembly and the Web Audio API, and the people at the W3C and the browser makers who built them.
SoundFont is a registered trademark of Creative Technology Ltd. Google AdSense is a trademark of Google LLC. All other names belong to their owners. The full licence texts are kept with the site's files. If your work is used here and the credit is wrong or missing, please get in touch through chaiwebs.co.uk and it will be put right.