discussion / Acoustics  / 23 September 2020

Distinguishing Noise from Sounds - a question of recognition?

Hi there,

Does anyone know of anyone researching into the identification of different noises, as opposed to distinct sounds, and can they point me at them please?

By "Noise" i mean multiple asynchronous multi-source sounds which equate to radio "Squelch", and by "Sounds" I mean single source audio output which can be identified (with practice).

We are working on identifying large flocks of birds who all tweet and flap asynchronously at www.withymbe.info

Many thanks,

Andrew.




This is also an issue in marine acoustics that I have worked on. To improve detection of the target sounds you need to develop classifiers for the soundscape i.e. your noise, and then build an adaptive response to that noise.  In the marine ultrasound context the noise may be sediment transport noise (huge and varied - many weak sounds), wave noise, shrimp noise (fewer very loud transients), boat sonars, acoustic modems, unknown biolgogical sources (diel chorusing is common and sources generally unknown) etc etc so this requires a layered approach to classifier development. ... but then you hit the problem of getting good generalization performance when you have too few soundscapes and many cannot be labelled.  So it is a limiting factor.

I'm guessing that the terrestrial problem is easier because sources are much more often known, but that's not a lot of help to you!

Thanks Nick,

In discussion here this issue of unknown classifications may be a widespread problem for supervised learning models? A colleague PhD candidate here is working on sound classification and mentioned the self same problem today. For example, "Cat", "Dog", "Everything else"? Is the "Everything else" state space (much) bigger than the two classifying state spaces?

One suggestion has been to take very small slices of "Squawk" of the flock of birds and use a distinct slice of frequencies from the spectrum thereof?

A problem remains however: how distinct is the combined squawk of a Quelea bird from other flocks of small birds?

Any advice on how to build specific classifiers would be most helpful?

Many thanks,

Andrew.

It sounds like you are referring to what is often called the "Cocktail Party" problem.  There has been quite a bit of research in this area, but it is a very hard problem.  I would start with the general literature on this topic as there isn't much in the conservation space. Check the hearing aid literature....picking out someone talking in a crowd of voices and sounds.  Humans are pretty amazing at this...machines not so much.  You might consider following up on some recent research at MIT:

https://blogs.nvidia.com/blog/2018/08/28/music-youtube-cocktail-party-problem-ai-artificial-intelligence-deep-learning/

For what its worth, it's a fascinating problem and has a lot of conservation applications.  As Nick mentioned, the biggest challenge is getting enough usable soundscapes.