MEL VS linear spectrograms for bioacoustics machine learning

Question

I don't have background in bioacoustics but working on a data-science project in bioacoustics.

I am working with animal vocalizations recorded at sampling rate of 250000.

Animals are bats, which are known to produce sounds in high frequency.

In background literature, I found the use of MEL spectrograms also for bats, and learnt that MEL spectrograms compress the range into log-scale.

But if so, I would expect infomation is lost for highest frequencies, and if so, I would expect worse results in using MEL.

Now, for the sake of benchmarking, I would like to reduce the number of variables of my study, and if other works used MEL, well there shoudl be a reason.

First goal : I am trying to identify vocal units, i.e. discretize vocalizations that seems (at human perception) kind of continuous screeches.

• Can you help understand pros and cons of MEL VS linear spectrograms , when dealing with animal vocalizations?

• Can you mention some good practices to keep in mind, considering the above goal ?

Answer 1

The most important is to set the parameters of the mel spectrogram appropriately. This means both the temporal resolution, the frequency range and frequency resolution. For example, with librosa.feature.melspectrogram this would be configured using the parameters n_fft, fmin/fmax and n_mels. If you have existing papers that use mel spectrograms with success on a similar task, use the parameters they selected as a starting point.

Linear (STFT) spectrograms have many more bins on the frequency axis. This is usually unnecessary (for most tasks), and the higher dimensionality makes it harder for a model to learn patterns from.