Onset tweaking

I’d like to use librosa onsets for separating transients for sampling. I noticed the peaks are always a little later than where I’d usually chop up a sound, so I wrote a little code to walk them back to the most recent local minima:

def tweak_onsets(onset_frames, o_env):
    tweaked = []
    for onset_frame in onset_frames:
        while True:
            if onset_frame == 0:
                break
            cur = o_env[onset_frame]
            prev = o_env[onset_frame - 1]
            if prev > cur:
                break
            onset_frame = onset_frame - 1
        tweaked.append(onset_frame)
    return np.array(tweaked)

Some onsets before:

And after:

It’s not exactly right, but for my application it works pretty well. I’m not sure that this is something librosa is designed for, and therefore not a bug per se, but I can imagine someone else wanting to use librosa to chop up sounds into transient samples, so I thought I’d leave this note here.

[bmcfee]

Compromise: factor into it’s own function, and optionally call it from onset_detect?

• onset_backtrack(events, energy) returns events (frame indices) rolled back to the most recent local min of energy
• onset_detect(..., backtrack=True, energy=None) calls above with energy=onset_envelope
• onset_detect(..., backtrack=True, energy=e) calls above with energy=e as a pre-computed energy curve e (eg, output of feature.rmse)

Sound reasonable?

[bmcfee]

Thanks @kylemcdonald for responding.

I’m thinking about the API here, and it’s not completely obvious how best to implement this.

One option is to add an energy argument to onset_detect. If energy is given, then backtracking to local min is enabled. Otherwise, we leave it as is. Note: we can’t directly compute energy from within onset_detect because it’s input may just be the onset envelope, and not the input signal.

Backtracking on the onset envelope itself is probably not correct, but maybe better than nothing? Should we make that an option as well?

Alternately, we could just make backtracking into a separate function that takes in onset times and an energy signal, and does the backtracking.

[Tagging @craffel for thoughts]