2023 03 29

After briefly walking through some more or less familiar Git functionalities with Dave they somehow arranged in my head more sensibly and created this repo.

Main changes compared to the version I had until now is that I made two notebooks for local and Colab execution, a separate branch for onset detection and opened up a few issues in the main.

I tried not to code again, but could not resist well. I implemented use of librosa.stream in the main branch. Now pretty big files (even around 1h) can be processed quite efficiently. Seems like the main bottlenecks were file loading and MFCC calculation which I both blame on librosa, which was even worse than the quick and dirty audio2numpy. This works very smoothly for evenly-sized segmentation of audio files, even with the most crude solution of block_size being set to 1, where each block is a single audio frame big files don't seem to make the script choke.

I plan to focus on for now, as onset detection will follow most of the same logic, with some caveats (programmer word) regarding the frame size.

Cool. onto other issues now.