Feature: Onsets

[tkell]

I suspect we can get this with librosa.onset.onset_detect

[Cortexelus]

Onsets make good segments, but the onset times themselves are not the best boundaries.

1. We want to find the nearest zero-crossing
   • less likely to get clicks when you concatenate
2. We want to set the segment boundary earlier to include the pre-onset head/attack.
   • less likely to hear next segment in the tail of the first
   • locate by backtracking to point of lower spectral flux
   • observe the various ways (from this IRCAM MuBu paper)
3. We may also want to include
   • time of max amplitude (originally in the echonest api)
   • good points to ping-pong loop this segment's samples for an Ableton-style time stretch. Perhaps a period of low spectral flux at least 100ms.

This is my aesthetic opinion from building concatenative synths and playing with librosa's onset_detection

[bmcfee]

@Cortexelus fair points. Have you seen the onset tweaking thread in the librosa issues?

[Cortexelus]

Yeah. A simple backtracking method like the one @kylemcdonald explains is what I was thinking. (But now this time reassignment stuff is new to me and I am exploring it.)

[Cortexelus]

It's useful to include both an onset time and a starting boundary. For example, in a rhythmic concat synth, you may want to start the segment early (pre-onset head) but align its onset time to the beat/pattern grid. This way a lengthy head doesn't mess with the pulse of the pattern (example: reverse cymbals), but you still hear the full head (overlapped with the previous segment's tail)

[Cortexelus]

I summon Tristan Jehan's thesis.

[tkell]

Closed with #95 !