I think we should raise a warning in here
|
gt_x = resample_audio(gt_x, fs, 16000) |
For many speaker embedding models like ECAPA-TDNN and RawNet3 if fs is less than 16kHz the speaker similarity is very bad.
E.g. on Fisher I observe ~0.65 vs using WavLM-SV that gets ~0.95
I think we should raise a warning in here
versa/versa/utterance_metrics/speaker.py
Line 58 in 55514f3
For many speaker embedding models like ECAPA-TDNN and RawNet3 if fs is less than 16kHz the speaker similarity is very bad.
E.g. on Fisher I observe ~0.65 vs using WavLM-SV that gets ~0.95