Embeddings dimensions

[carlos-aguayo]

If I get the embeddings using speaker-diarization

p = Pipeline.from_pretrained("pyannote/speaker-diarization-3.0", use_auth_token=True)
diarization, embeddings = p(file, return_embeddings=True)

Then the dimension of embeddings is 256.

If I get the embeddings using:

model = Model.from_pretrained("pyannote/embedding", use_auth_token=True)
inference = Inference(model, window="whole")
embedding = inference.crop(file, segment)

Then the dimension of the embedding is 512

What can I do to ensure that I get the same dimension on both invocations?

[hbredin]

pyannote/speaker-diarization-3.0 does not rely on pyannote/embedding internally.
Instead, it relies on hbredin/wespeaker-voxceleb-resnet34-LM, hence the different dimension.