Advice on the best methods for quantifying high-medium SNR bird sounds

Question

Related to my recent post, I am working on finding a way to 1) diagnose the presence of two kinds of specific bird sounds throughout my hundreds of hours of audio data, and perhaps to 2) analyze the success of my detection routine. I am aware of R package ohun as a great option for both of these, but due to problems and the lack of information about it online, I am considering using a different process for sound detection.

I have access Raven 1.6 Pro but have not yet used it for sound detection - wondering if it is a good option and if there is a recommended approach? The sounds I’m trying to detect have a relatively high signal-to-noise ratio and are quite easily visualized in my spectrograms. For each of the sites at which I recorded audio of the birds (so for each of my sound files), I will have a specific threshold at which these sounds are made by the relevant focal birds and will need to be "included", while other sounds of this type that are the same sound, but quieter, will need to be discounted. Again, ohun provides great ways to do this, but I might need a backup plan. Feel free to ask questions since I am not sure what information about my data is relevant to my sound analysis decisions.

I would love to hear from folks who 1) have suggestions for how I might quantify high/medium SNR sounds from many hours of data or those who 2) have used ohun successfully and would be willing to chat about troubleshooting.

Answer 1

Assuming that the expression "quantifying high/medium SNR signals" of "two kind of specific bird sounds" implies that the signals have unique characteristics that allow discrimination between the sound types and between the signals of interest and the rest of the world, the most promising method are matched filters See here.

Practically, you correlate (filter) your data with a snippet of the time-reversed time series of the sound of interest. The resulting vector will peak when there is a good match between data and signal of interest. The processing gain is proportional to the time-bandwidth product of the signal, which means, it does not work so good for constant tonals (little bandwidth) or impulses (little time duration).

If the data are too long for efficient correlation, then there is the "spectrogram correlation", which takes a fraction of the spectrogram as replica to processes the data in the spectral domain. Keyword for searching the literature is "spectrogram correlation". It is used frequently in automated underwater sound detection (dolphin whistles and baleen whale calls). AFAIK, Raven should have the spectrogram correlation implemented.

In recent years, spectrogram-based detections are frequently the base of Convolutional Neural Networks, so it is worth looking into it.

Re ohun or other software packages, however, I'm of little help, and someone else must step in.