How does SSPS evaluate on VoxCeleb1-O dataset?

[hoangducnhatminh]

I see the paper you mention that you evaluate on the full-length utterances in section 3.4, but the code did not say the same. Can you confirm?

Thank you for this awesome project.

[theolepage]

Hi,

During training (using the train*.py scripts), evaluation is performed on 4-second segments (64000 samples) to speed up computation. In contrast, full-length utterances are used during evaluation (with the evaluate*.py scripts), as specified in the SSPS paper and reflected in the configurations:

sslsv/models/ssps/voxceleb2/simclr_e-ecapa/ssps_kmeans_25k_uni-1/config.yml

Line 23 in 15afa50

frame_length: null

Thanks!