The released model corresponds to the paper 'CoarseSoundNet: Building a reliable model for ecological soundscape analysis' (paper link).
The model tags audio files to one or more of the three coarse soundscape components Anthropophony, Biophony, or Geophony.
It also predicts silence, but we recommend to ignore this class and assume silence only when none of the three classes was predicted.
Installation
To use the model, you have to install autrainer, e.g. via pip:
, where -w is the window size in seconds, -s is the step size in seconds and -sr is the sampling rate.
For other possible inference settings and all usable parameters, please have a look at the autrainer documentation.
However, the above settings are recommended.
Caution! The sampling rate needs to be kept as given above (48000), in order to process the files correctly.
Training
We conducted our training via the autrainer library by Rampp et al. (2024), which enables reproducible training.
Dataset
We used Edansa-2019 (Çoban, E. B., et al., 2022) together with own annotated data from the University of Freiburg. For more details, please be referred to our paper.
Features
The audio recordings were resampled to 48kHz. We then used the CLAP-FeatureExtractor ('laion/clap-htsat-unfused') from Wu et al. (2023).
For more information, please check our github repository (will be made public soon).
Training process
The model has been trained for 30 epochs. At the end of each epoch, the model was evaluated on our validation set.
We release the state that achieved the best performance on this validation set.
All training hyperparameters can be found inside conf/config.yaml inside the model folder.
Evaluation
We evaluated our model on both the test set from Edansa-2019 as well as a strongly annotated dataset (BEsound) from ecologists at the University of Freiburg.
The audio recordings of the specifically annotated BEsound data are 60s long.
In order to evaluate the model, we check the full recording for whether there is one (or more) of the three target classes predicted / present.
On Edansa-test we achieve a macro F1-score of .892 and on BEsound we achieve a macro F1-score of .758.
For more information please be referred to our paper.
Acknowledgments
Please acknowledge the work which produced the original model. We would appreciate an acknowledgment to autrainer.
Bibliography
Gebhard, A., et al. (2026) "CoarseSoundNet: Building a reliable model for ecological soundscape analysis." arXiv preprint arXiv:2605.21143 (2026).
Rampp, S., et al. (2024). Autrainer: A modular and extensible deep learning toolkit for computer audition tasks. arXiv preprint arXiv:2412.11943.
Çoban, E. B., et al. (2022). Edansa-2019: The ecoacoustic dataset from arctic north slope alaska. In Workshop on the detection and classification of acoustic scenes and events.
