alecristia
commented
2 years ago In the Vanuatu dataset, some children have recordings that start in the night time. For instance, 1_2016-09-26 starts at 20:30. Often, this happens because the recording is long and it got re-started (for tech reasons or whatever). That same child has a second session starting about 7h after the first one, at about 6:30am, so in a way one could think they belong to the same session and/or that the first one should be disregarded when estimating day-time vocal activity.
We can go about this in at least three ways:
- Calculate metrics in a time range
For instance, we may want to only look at 6am to 6pm. So for the example child mentioned above, the first session would be ignored and the second included. Another child may have a session that starts at eg 2pm and ends at 10pm -- then only the section of the audio corresponding to 2-6pm would be considered.
Note that "for all the recordings that are concatenations of the original clips, start_time_accuracy = hour (because we are not sure there is no gap within the recordings, so we cannot infer clocktime at every point) for the others, start_time_accuracy = minute until one audio is shorter than the usual 4 hours due to the USB splitting the audio); then, start_time_accuracy = hour for all consecutive audios within that session following any audio shorter than 4 hours".
In some datasets (Tsi 2017) we don't have start times at all, but only whether the rec was launched in the AM or PM.
- Filter recordings based on start time only
This would work including in datasets like Tsi, but it's (a) a waste of data, since some recs started at eg 2am may nonetheless contain useful info (from 6am on); and (b) noisy, since some recs started at eg 5pm are essentially night-time ones
- Leave metrics at the session level, then use weights/other procedure in stats to account for variance.
For now, in Vanuatu, I'll do the last one.
But option 1 does seem the best!
lucasgautheron
Is your feature request related to a problem? Please describe.
Usually we are only interested in daytime and some recordings may contain a lot of night time. There should be a way to recover the timestamps of the recordings that match a (clock_start_time, clock_end_time) range specified by the user.
Describe the solution you'd like
We should discuss the alternatives here and try to see the bigger picture (e.g. sometimes, we don't want to throw away some of the data, but we can to aggregate statistics e.g. hour by hour, and maybe we can find a way to approach these two tasks similarly)