Using spkrec-xvect-voxceleb For Verification (All Comparisons Scores Are High)

[MattyB95]

I sought to perform speaker verification tasks using the default spkrec-xvect-voxceleb and spkrec-ecapa-voxceleb models. The ECAPA model works as expected. However, the xvect scores high no matter the comparison.

For example:

from speechbrain.pretrained import SpeakerRecognition

verification = SpeakerRecognition.from_hparams(source="speechbrain/spkrec-xvect-voxceleb",
                                               savedir="pretrained_models/spkrec-xvect-voxceleb")
score, prediction = verification.verify_files('id10001\\00001.wav', 'id10002\\00003.wav')  # Different Speakers
print(score, prediction)
score, prediction = verification.verify_files('id10001\\00001.wav', 'id10001\\00002.wav')  # Same Speaker
print(score, prediction)

Output:
tensor([0.9933]) tensor([True])
tensor([0.9982]) tensor([True])

I would expect the different speakers to score much lower than the same speaker, but regardless of which audio files I compare, all scores are > 0.99. I need to do something entirely differently when using xvect for verification.

[maxpain]

The same problem