I've noticed too many conversations lack balance.
** Let me know if you're interested in helping with this **
What it does
It is a small, unobtrusive box that listens to conversations using a microphone and tries to identify distinct voices and the genders of these voices. Companies could install it in meeting rooms, and schools could install it in classrooms. If it detects a very skewed discussion, it lets the users know through a gentle visual reminder.
It won't be very accurate, since the machine learning task here is very difficult and even humans aren't perfect at it. But if:
90% of the voices are from 3 distinct people or fewer over a rolling 30 minute period
75% of the voices are from only one gender over a rolling 30 minute period
10 incidents of interruption occur in a 30 minute period
it will start to gently glow red in order to prod the participants to have a more balanced discussion.
Again, I don't expect it to be very accurate, since it may not even be possible to perform this task perfectly. But if the thresholds are fairly high, it should be a useful tool in practice.
The thresholds will of course be adjustable for different environments. A web interface will display the analysis results in real time.
How I built it
I'm still working on this. By the end of the hackathon, I hope to have gender identification and the glowing red LED working. (Open source code is already available for gender identification of voices.) I think I can also get a primitive Web display done.
Counting distinct voices and incidents of interruption are equally important, but they are more challenging and might take more time to implement.
I've done audio processing projects before, including the SoundScout airplane noise detector.