This converts a WAV file recording into another format (MIDI, flopkestra bytecode, etc.)
The input is a WAV audio recording. It is partitioned into windows (by default, each window is 2^12 samples long, about 0.09s for a 44kHz sample rate). The Discrete Fourier Transform is applied to each window to determine which frequency is dominant in this window. Each of these frequencies are rounded to the nearest frequency that corresponds to a standard MIDI note.
From these, a list of pairs (n, d) is created, where n is the number of
the MIDI note corresponding to the frequency and d is the duration of that
note in seconds.
The stream of pairs (note, duration) created by the previous step can be
transformed by filters. For example, one might want to weed out all notes that
have a too short duration, or bump all the note one octave up, or else to get
the duration of the notes in a different unit than seconds.
The filtered output is then given to one or more writers to be converted into another format. At the moment, there are writers for MIDI files, flopkestra bytecode and JSON.
The only external dependency is numpy, which is used for the discrete Fourier transform computations.
This was built by Rafik Draoui to make his floppy drive orchestra play tunes he has in his head without having to figure out how to write them down in MIDI or flopkestra bytecode by hand.