Some pretty belated benchmarks have proven there are some rather fundamental issues with the clustering quality. Report writing actually provided some insight into this matter, with some problem solving done during writing the problem solving section.

As should be expected with a pipeline at this scale, the bugs have a tendency to cascade EXTREMELY. A lot of the model output tended to be semi-random, even after fixing the architecture. Again, writing the report made it seem like the issue just requires backtracking into how the classification works and its accuracy, then further back into feature extraction, eventually showing that a lot of issues might have been simply caused by the segmentation.

I have visualized the slicing, but somehow wasn't alarmed enough by it being obviously out of sync. Quick fixes helped realize the clustering improves (finally), and, further tests showed that the gen~ patch actually does introduce error in playing back the grains.

No time to debug the gen~ synth, so I simply rolled back to the stock granular synth implementation by Federico Foderaro, maybe with an A/D envelope control available alongside the current granular-synthesis style windows.

These new changes are atm included in the debug branch only, will wait for the main cleanup time today or tomorrow with merging them into main.

Now, I should stop having fun testing out the somewhat working (or being slightly more convincing) project and keep on writing the report!