predicted cluster number always be 2 about spectralcluster

[wq2012]

What you are using is the vanilla spectral clustering without any refinement. Refinement is critical for more accurate estimation of speaker count. You can use these configs as examples: https://github.com/wq2012/SpectralCluster/blob/master/spectralcluster/configs.py

[kli017]

Thanks for the remind. I tried the turntodiarize cluster in the configs. I noticed it need a constrain matirx computed by turn score vector while doing predict. However I do not have that score. Instead I computed the distance matirx of the embddings and set the CL and ML according to the distance. But I still just get 2 speaker as result. If I use the distance matrix direclty I can get multiple speaker, however I found the result does not good. I was wondering how should I get a constrain matirx properly?
Thanks.