Comment by hamza_q_
2 months ago
Thanks for checking it out!
Yeah unfortunately, since the diarization is acoustic features based, it really does require high recorded voice fidelity/quality to get the best results. However, I just added another knob to the Diarizer class called mer_cos, which controls the speaker merging threshold. The default is 0.875, so perhaps try lowering to 0.8. That should help.
I'll also get around to adding a oracle/min/max speakers feature at some point, for cases where you know the exact number of speakers ahead of time, or wanna set upper/lower bounds. Gotten busy with another project, so haven't done it yet. PR's welcome though! haha
Thanks, `mer_cos` definitely gets me closer. I appreciate that. Yeah, I was thinking providing a param for the expected number of speakers would be nice. I'll check out the codebase and see if that's something I can contribute :).
Yeah would love contributions! Here's a brief overview of how I think it can be done:
Senko has two clustering types, (1) spectral for audio < 20 mins in length, and (2) UMAP+HDBSCAN for >= 20 mins. In the clustering code, spectral actually already supports orcale/min/max speakers, but UMAP+HDBSCAN doesn't. However, someone forked Senko and added min/max speakers to that here (for oracle, I guess min = max): https://github.com/DedZago/senko/commit/c33812ae185a5cd420f2...
So I think all that's required is basically just testing this thoroughly to make sure it doesn't introduce any regressions in clustering quality. And then just wiring the oracle/min/max parameters to the Diarizer class, or diarize() func.