Part of the tutorial from my blog poissonisfish.
The MP3 files and the optional metadata CSV file are part of the xeno-canto collection and hosted in the public Kaggle dataset Bird songs from Europe (xeno-canto). Rights to use and share them are listed under the Terms of Use page from xeno-canto.
- Download, unzip and move
mp3/from the Kaggle dataset to your GitHub repo clone - Make sure all packages listed on top of
0-prepAudio.Rand1-classAudio.Rare installed - Execute
0-prepAudio.R - Restart R (recommended) and execute
1-classAudio.R
The analysis was conducted using an iMac Pro machine with AMD graphics cards and 16 threads. Considering Tensorflow GPU support is exclusive to NVIDIA®, the much less restrictive plaidML framework was used instead, deployed from a Conda environment called plaidml. Alternative setups including TensorFlow backends (GPU or CPU-supported) can be easily accommodated.
Also, consider changing ncores in 0-prepAudio.R as suits your machine specs the best.
To further improve reproducibility, the Keras model object model.h5 is included to validate the reported accuracy. Additionally, sessionInfo carries detailed information of the package versions and hardware configuration.
Feel free to report any issues with reproducing the analysis described in the blog.
I want to thank the xeno-canto community for their work in collecting, documenting and sharing this information.