I play with Magenta models for generating music, strictly for my own amusement. You can find much better demos and sample code on the
magenta website.
In particular, I use the onsets and frames model, but
the piano scribe demo does it better.
You need npm.
npm install
npm run build
node dist/index.js
This will bring up a web server listening on port 3000. You can connect to it locally. Depending on how restrictive your browser settings are, it is possible that the application cannot get access to your microphone. A freshly installed Firefox or Chromium browser works for me (on OpenSuSE).
If you want to run this on a web server and access it over the internet (though I honestly don't know why you would), you need use https.
Click on the microphone icon to start recording sounds. Whistling works well. Your browser should ask your for permission to use the microphone. Whistle (or record some other sound), then click on the microphone again to stop recording. The application will stop recording by itself after 10 seconds.
The page will take a few seconds to transcribe your sounds to midi,then you can play them. After that, you can use the "Compose" button (with the note on it) to ask the computer for a composition inspired by your input. You can select a style via the radio buttons. The output is nondeterministic, so you may want to try several times.
The composer will pre-process your inputs and the result of that is also shown. It tends to be a little easier on the ears than the midi from the original, though maybe that is just my bad whistling.
Note that most of the heavy lifting is done in the browser, hence the long load time. There's some background and possible alternatives here: https://www.nephometrics.ch/2019/06/serving-magenta/, I just haven't got around to implementing them yet.