Design dataset features

[kureta]

How are we going to represent each music imported from midi files?

Initial suggestion:

• interval: In semitones. Interval of initial note is zero.
• delta time: Time between last notes onset and this notes onset. This way we can easily represent arpeggios and chords.
• duration: Maybe in seconds, maybe in beats, or as a fraction of the duration of one measure. If in beats or measures, we may also need tempo and/or time signature.
• velocity: Very simple, just normalize midi velocity to [0, 1]
• register: This might be tricky. Normalize the range of the piano to [0, 1] and add uniform noise with a certain width (maybe half or quarter of an octave) so that the network cannot learn exact pitches, just a general register. This will be necessary since we are using intervals instead of pitches, we might go outside the register.

[kureta]

We might need to use plain old seconds for duration. Otherwise, during live interaction, players have to follow a fixed tempo and time signature. #1