Maximum sample rate for CNN training

[Maxime-Bru]

Hello,

By default, when training a custom CNN, without modifying the preprocessor, the native sample rate of the audio files is taken and therefore the spectrograms images are comprised between 0 and fs/2, am I correct? There is no resampling of band-pass filtering by default right? Does this hold even for ultrasonic recordings where fs can be up to 256kHz or even higher?

Thank you,
Maxime

[sammlapp]

Hi Maxime,

The current default behavior is no audio resampling, but cropping spectrograms to the frequency range 0-11,250 Hz

opensoundscape/opensoundscape/preprocess/preprocessors.py

Line 548 in 2f14a07

Spectrogram.bandpass, min_f=0, max_f=11025, out_of_bounds_ok=False

The reasoning for this is that audio resampling is computationally expensive, but I have considered changing the default behavior since users may assume audio will be resampled to a constant rate.

You can set the preprocessor to resample audio:

model=CNN(...)
model.preprocessor.pipeline.load_audio.set(sample_rate=32000)

and/or can modify or disable spectrogram bandpass similarly:

model.preprocessor.pipeline.bandpass.set(min_f=0, max_f=16000)
# or  turn off spectrogram cropping
model.preprocessor.pipeline.bandpass.bypass = True

See the preprocessing tutorial for more detailed work-through of customizing preprocessing, and let me know if you have further questions

Best,
Sam

[Maxime-Bru]

Thank you Sam,
I changed the bandpass and it suits my needs!

Maxime