Explicit control over sound event overlap

[justinsalamon]

Certain scenarios require limiting the amount of overlapping sound events (or prohibiting it altogether). Right now there is no way to explicitly control for sound overlap.

Proposed solution: add a max_overlap kwarg to generate(), which by default is set to None, meaning any overlap is allowed. If set to 1, it basically means no overlap is allowed, 2 means 2 overlapping sounds, etc. 0 would be an invalid value.

(This would address #62)

[pseeth]

To tag on a bit, a min_overlap could also be nice, to make sure that generated scenes aren't trivial for source separation networks to solve, if using Scaper for the purposes of training separation networks.

[justinsalamon]

Sure! The logic for the two is quite different though - max_overlap just needs to find a place to locate each event such that the constraint isn't violated. min_overlap would basically require adding more and more events until the minimum overlap is reached everywhere. Will probably implement max_overlap first and then min_overlap in a separate PR (once pending PRs are finally merged.. sheesh! this fall hopefully!)

[sreenivasaupadhyaya]

Hi @pseeth @justinsalamon , I was wondering if the max_overlap is already implemented in any branc/version of Scaper already?

If not could you guide me to get it done? Thanks !

[justinsalamon]

@sreenivasaupadhyaya it’s not implemented yet. The feature requires a couple of delicate design decisions. We’re currently blocked on higher priorities, but will try to get to this soon, thanks

[Kovalt0]

@sreenivasaupadhyaya it’s not implemented yet. The feature requires a couple of delicate design decisions. We’re currently blocked on higher priorities, but will try to get to this soon, thanks

@justinsalamon is this tool already implemented? I am trying to create thousands of audio samples in which they have from 1 to 3 events without overlapping. I managed to implement a for loop that cover the "for _ in range(n_events):" part and it kind of works but still not what I want to do. In the documentation I also couldn't find a way to make it work by setting the onset of each event. I tried with a list, but it stills overlap the sounds.

In this example I created soundscapes with three events, the implementation for creating 1 and 2 events samples is commented.

# """ 1 event"""
# time1 = round(random.uniform(0, 14), 1)
# time_plus = [time1]
# """ """

# """ 2 events """
# time1 = round(random.uniform(0, 4), 1)
# time2 = round(random.uniform(10, 14), 1)
# time_plus = [time1, time2]
# """ """

"""3 events"""
time1 = round(random.uniform(0, 4), 1)
time2 = round(random.uniform(10, 14), 1)
time3 = round(random.uniform(20, 24), 1)
time_plus = [time1, time2, time3]
""" """

# time_plus = [time1, time2, time3]
for event_time in time_plus:
    print('event time is: ', event_time)

    # add random number of foreground events
    n_events = np.random.randint(min_events, max_events+1)
    for _ in range(n_events):
        sc.add_event(label=('choose', []),
                 source_file=('choose', []),
                 source_time=(source_time_dist, source_time),
                 # event_time=(event_time_dist, event_time_mean, event_time_std, event_time_min, event_time_max),
                 # event_time=(event_time_dist, event_time_min, event_time_max),
                 event_time=(event_time_dist, event_time),
                 # event_duration=(event_duration_dist, event_duration_min, event_duration_max),
                 event_duration=(event_duration_dist, event_duration),
                 snr=(snr_dist, snr_min, snr_max),
                 pitch_shift= None, #it accepts None value
                 time_stretch= None) #it accepts None value
pass

Thanks in advance.