Should Chirpity show the whole recorded frequency range in the spectrogram?

[ceperman]

I've made some recordings at 48kHz sample rate which gives a frequency range up to at least 20kHz, but the spectrogram only shows up to ~12kHz. Is this correct? A spectrogram from Audacity shows bird sounds in these recordings well up to 17kHz. (File attached; Audacity at top)

We've had a discussion about the accuracy of detections based on different sample rates, and the Audacity spectrogram clearly shows information exists in the 12kHz-17kHz range, information that would be missed in recordings at 24kHz (which I've been using). Do we know if BirdNET uses information in this range? It would certainly be a good reason for recording at 48kHz if it does.

[Vollpflock]

The BirdNET model does not use frequencies above 15KHz. As far as I know, what you see in this spectrogram are just harmonics or overtones.

[Mattk70]

Yes, that is correct, the spectrogram view is capped at 12kHz. BirdNET is sensitive to sounds up to 15kHz and Chirpity sends it audio with a sample rate of 48kHz regardless of the source sample rate of your audio. Obviously, if your audio sample rate is 24kHz, there will be no information above 12kHz, so for BirdNET analysis it will be better to use audio with a sample rate of at least 30kHz. 48kHz would be the ideal rate.

[DD4WH]

We have to distinguish between:

1. frequency span that is sent to BirdNET and the one that is used by BirdNET
2. the frequency span that contains useful information for manual inspection of the bird song spectrograms and for distinguishing bird species

1.) the BirdNET model uses only frequencies up to 15kHz (as has already been correctly stated). Internally, BirdNET uses 48ksps sample rate. If you use a differing sample rate, your WAV files have to be resampled which takes some time and could potentially lead to artifacts. However the latter point is very minor and should not result in fundamentally different results in the classification.

2.) Birds (at least in Europe) mainly use frequencies up to 10kHz. All higher frequencies are non-relevant for the ID. So, for a manual inspection of spectrograms for ID purposes, we need a span from 100Hz to 10kHz. All higher frequencies are non-relevant for the ID. In order to have a max freq resolution in a given spectrogram, we should scale the spectrogram for this freq span. For example, have a look into Stanislaw Wrozas book "Identifying migratory bird sounds in Britain and Europe": the spectrograms there show freqs between 0 and 11kHz.

So, in conclusion in order to answer this threads question "Should Chirpity show the whole recorded frequency range in the spectrogram?":

No, Chirpity should not show frequencies above 12kHz in the spectrogram (at least if we are talking about bird distinction)

I would further recommend using 48ksps sample rate in order to give BirdNET the whole spectrum it uses (up to 15kHz) AND to simultaneously save time and processing power, because no resampling is needed.

P.S.: your recording, ceperman, looks quite strange . . . the harmonics seem to indicate an overload of the ADC of your recorder or something else. What is your recorder model ? Was the bird very close ? Was it a field recording or did you record from a playback with a low quality speaker (eg. a smartphone)? The latter would explain the strange looking spectrogram

[Vollpflock]

@DD4WH
I know, off topic, but maybe you could help me with that: Are harmonics (or overtones, whatever they are called) always irrelevant for bird sound identification? Also, really great to hear from you again! I’ve actually been saving up some audio hardware questions, hoping I’d get the chance to ask you one day

[ceperman]

@DD4WH
If I fully understand what you're saying, it seems that there is nothing really to be gained from recording above 32ksps. All the information used by BirdNET will be captured. The only issue is the resampling done by BirdNET (or Chirpity in our case) and any extra processing is borne by my computer. As long as detection accuracy is essentially unaffected, that's OK by me.

You also say:

P.S.: your recording, ceperman, looks quite strange . . . the harmonics seem to indicate an overload of the ADC of your recorder or something else. What is your recorder model ? Was the bird very close ? Was it a field recording or did you record from a playback with a low quality speaker (eg. a smartphone)? The latter would explain the strange looking spectrogram.

This is perhaps not surprising as this a home-made recorder, based around an ESP32 microprocessor and and I2S INMP441 microphone. The mic does seem to be very sensitive, and within the code I do try to reduce the gain, but this doesn't affect the fact that signal can already be distorted by close/loud sounds before it's delivered to the processor. Is what you see going to affect its usefulness?

As an ex-developer and sometimes electronics tinkerer, I built it because I thought I could, and at £20 it turned out cheaper than buying one!

[DD4WH]

@Vollpflock No, harmonics can be quite relevant to bird ID. Look at the spectrograms of the calls of willow tit or look here for laughing gull song that contains a lot of harmonics (https://becausebirds.com/visualize-bird-songs-with-spectrograms/). However, harmonics also occur when the mic, ADC, preamp is being overloaded, this mainly accentuates the uneven harmonics.

@ceperman Very interested in your recorder! Do you have a description and your code on github? Would be very interested. Hmm, not sure whats happening there in your recorder, but I would not trust the recorded freqs above 10kHz. Also, the datasheet of the INMP441 indicates that freqs above 10kHz are not recorded accurately, if you look at fig. 4 for example. I use ICS43434 as a very good I2S mic, have a look at Adafruit, they have breakout boards with that mic.

[ceperman]

@ceperman Very interested in your recorder! Do you have a description and your code on github? Would be very interested. Hmm, not sure whats happening there in your recorder, but I would not trust the recorded freqs above 10kHz. Also, the datasheet of the INMP441 indicates that freqs above 10kHz are not recorded accurately, if you look at fig. 4 for example. I use ICS43434 as a very good I2S mic, have a look at Adafruit, they have breakout boards with that mic.

Happy to provide more info on my recorder, but best not pollute the Chirpity discussion group by doing it here. I have written about it on a blog I have: https://ceperman.blogspot.com/2025/02/an-esp32-based-bird-sound-recorder.html

I did have problems with harmonics during development and thought I'd got rid of them. Perhaps I need to recheck.