Help in making a Birdvocals recognition system

[TomChiron]

Using a Teensy 4.1 just as a Soundcard to get one microphone signal into a Raspberry Pi looks like overkill to me - especially considering the goal of that great project idea to have a system as cheap as possible. I have no experience with Raspberries, but I am sure there would be other solutions which fulfill the need to have it small. There should be boards/ shields which even would make the building of the system easiert than dealing with a Teensy.

From the datasheet the AOM5024 seems to have a really good signal to noise ratio. The 80dB sound a bit too good to be true. Point is that there are many other small microphone capsules even in a range of 60dB S/N-ratio in the datasheet which should be sufficient. For example, you can make excellent recordings with the Panasonic WM61A which is not longer produced. According to Digikey a substitute would be the POM-3535L-3-R.

From the signal quality, i.e. mainly looking at the frequency response, all small microphone diagrams give a flat = linear response because that is physics.

I expect it to be more challenging to have a good mechanical protection from moisture and to have something to supress wind noise. A huge advantage of omnidirectional microphones is here that they are far less sensitive to that.
Another thing could be structure borne noise. When the housing of the system is mountend in a way that it can vibrate, move, or can take noise over from the mounted structure that could be a problem. But I would not expect too many problems here.

 

[YvesD]

EVERY single BirdNET detection has to be carefully checked manually by visual inspection of the spectrogram and listening BEFORE submitting to xenocanto or ebird.
Please do not implement any "automatic" batch upload possibility of unchecked recordings !

Hi DD4WH,

That is NOT THE IDEA to do anyway !!!

What would be the fun in observing Migration if it's automated (and run by AI)...

It is more like a "add-on" helper on the station. Think of a fieldrecorder with a screen...

If "rare birds" are "detected", at least we have some audio of it to examine afterwards !!!
We have a strong community of "NocMig" and "VisMig" in my region, we have some exellent experts on birdsounds too.

So no, that won't happen !

 

[YvesD]

just to be sure ...

As this forum is not a chatbox...
Wouldn't it be better to continue the non-technical conversations (don't want to upset the Admin angry !) in a other way?
Over Github? Any other way?

Github repo BirdNET-Pi-MigCount

The technical things could stay here to benefit everyone...

Grtz,
Yves

 

[DD4WH]

From the datasheet the AOM5024 seems to have a really good signal to noise ratio. The 80dB sound a bit too good to be true. Point is that there are many other small microphone capsules even in a range of 60dB S/N-ratio in the datasheet which should be sufficient. For example, you can make excellent recordings with the Panasonic WM61A which is not longer produced. According to Digikey a substitute would be the POM-3535L-3-R.

You are right for loud Signals. But bird songs/calls tend to be faint signals far away. And for this, you would want the highest sensitivity and highest SNR. And, believe it or not, the AOM5024 and EM272 do really have these specs. I have used many of them for years and they are the best you can get for this application. Every vendor of expensive bird mic equipment uses them .
The WM61 is so extremely noisy, that it's useless for field recording.
I really recommend you test out the AOM5024, its really cheap and very high performance for field recording, you will not regret it.

I expect it to be more challenging to have a good mechanical protection from moisture and to have something to supress wind noise.

No moisture protection needed, they are robust enough, I have used mine without any protection for many years now. A foam windshield is enough for wind noise protection, because in heavy wind situations birds do not sing/call any more anyway and noise by trees is much too high in those situations.

 

[DD4WH]

That is NOT THE IDEA to do anyway

Thanks for clarification and sorry that I misinterpreted your aim!

 

[cebersp]

About the highpass filter: As long as there is no clipping at any point of the analog stage it is more useful to have a filter implented digitally in software. This way you are far more flexible to choose the exact type of filter with all its parameters.

Just to clarify my question about High Pass Filter:
I am doing a lot of sound processing for guitar music sounds. I that field highest amplitudes come from low frequencies. So, if you don't want to have any clipping or only moderate clipping, you have to tune gain for those low frequencies. It is quiet common, to put a high pass filter early into the signal chain, so gain can be higher and higher frequencies can be handled with better resolution. For this use of the high pass, it is useful to be before the ADC. - This was my background to ask the question here.
Also, if I have a look at my Birdnet recordings in audacity, the signal level of most bird sounds is very low. At the moment though, I am using a very small headset microphone, which I happened to have. The hope is, that the setup should record small birds, which are 50m away. We will see, if a bigger condenser microphone helps. I have ordered a condenser USB-micro and also some electret capsules.

The schematics, that I have posted in #17 has the feature to do "soft" clipping. This means, that the addition of harmonics does start moderately which gives some sort of compressor effect. Humans can normally only hear distortion >0,5%, so for them this type of compression works to a fair degree. I am not sure, what the Birdnet AI model does with "moderate" distortion. Another possibility for higher sensivity would be to use a "real" compressor (Automatic gain). One, that I like, is called Orange Squeezer: https://www.musikding.de/docs/musikding/squeezer/squeezerschalt.pdf These guitar effects are typically designed to work with 9V supply though.

What I wonder is, if the whole sound recognition could be done with Teensy. Probably the global birds model is far to big, but there is no benefit to have all these 6000 species in RAM. I have no idea how to generate a local reduced model. Of course this would benefit, if there was a Teensy variant with WLAN.....