It is difficult to find interesting data in a large corpus.

[kissake]

Given that someone is likely to sleep for 8 hours at a time, and reviewing 8 hours of data (audio recordings) in 10 minute increments is ... a lot (could take more than 8 hours), it is important to facilitate review.

[kissake]

Possible approaches to mitigate this issue include:

• Providing a facility to annotate the data easily (ideally without getting up)
  • Wireless connection (e.g. web page accessible from a cell phone), or
  • An audible signal that can be identified on an audio recording automatically (e.g. a DTMF tone generated by a phone dialer on a cell phone, or a fingertip rubbed on the microphone for maximum loudness).
• Automatically extracting or highlighting audio data that is interesting. Example criteria:
  • Louder than median loudness.
  • Occurs within some threshold before or after an interesting event (to capture audio that is unique in tone rather than level/loudness)
• Other?

[kissake]

Another tool to simplify locating useful information is to combine sequential segments of data into larger batches so that they can be processed at once. If other indicators can be retained (e.g. relative timestamps?), that would be ideal. Note that at this time the audio recording is explicitly cut early, with ~5 seconds of "grace", to avoid the next cron started recording from failing because the previous recording still holds the lock on the audio devices. This will cause some skew that will (for large number of sequential recordings) result in timestamps not being maintained if the recordings are simply concatenated.

[kissake]

A strategy for mitigation:

• Use concatenate-ffmpeg.sh to combine mp3s into a wav. (note the caveat above w/r/t 5 seconds of grace).
• Use Audacity to load the resulting WAV.
• Select all
• Amplify to almost clipping (default amplify amount)
• Select the quietest noise, use noise reduction profile.
• Select all, use noise reduction
• Select Analyze - Sound Finder, choose 8, 5, 1, 1, 0. If there are too few sounds detected, increase the first number by 4 and repeat. If there are too many, decrease the first number by 4 and repeat.
• You will want to remove the label track (below your existing track) before repeating this last step.
• Click on the number associated with each label to highlight the track in question. Play will play that section one time through.

The ideal would be to extract this data along with timestamp information into individual files, I think?