This code has been highly optimized to our lab-infrastructure and thus most probably has to be modified. The following parts are probably the most noteworthy regarding this:
- Data Generator
- Evaluation suite
- Setup
Since audios vary in shape, it gets difficult rather quickly to store them efficiently in a hierarchical data format like hdf5, zarr, etc. Furthermore, we also draw frequency vectors from arbitrary locations in an audio for our study. Those two factors made storing the audios in a "beautiful" way almost impossible. We had to perform those modifications as our storage at the time was a fileserver which would've slowed training and testing down by factor 100+.