| --- |
| license: cc-by-nc-sa-4.0 |
| pipeline_tag: audio-classification |
| tags: |
| - autrainer |
| - audio |
| - ecoacoustic-tagging |
| - HearTheSpecies |
| - ecoacoustics |
| --- |
| |
| # CoarseSoundNet |
|
|
| The released model corresponds to the paper 'CoarseSoundNet: Building a reliable model for ecological soundscape analysis' ([paper link](https://arxiv.org/abs/2605.21143)). |
| 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: |
|
|
| ``` |
| pip install autrainer |
| ``` |
|
|
| This model has been trained and tested with autrainer version `0.6.0`. |
| For more information about autrainer, please refer to: https://autrainer.github.io/autrainer/index.html |
|
|
| # Usage |
|
|
| The model can be applied on all wav files present in a folder (`<data-root>`) and stored in another folder (`<output-root>`): |
|
|
| ``` |
| autrainer inference hf:HearTheSpecies/CoarseSoundNet -r <data-root> <output-root> -w 10 -s 10 -sr 48000 |
| ``` |
| , 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. |
|
|
