Use an own trained SpeechBrain model

[matakanobu]

Hi, I am trying pyannote2.0 with speechbrain embeddings.

In order to use an own trained speechbrain model, the top directory of SpeechBrain model must be named "speechbrain".

pyannote-audio/pyannote/audio/pipelines/speaker_verification.py

Lines 268 to 271 in bbd5513

	if isinstance(embedding, str) and embedding.split("/")[0] == "speechbrain":
	return SpeechBrainPretrainedSpeakerEmbedding(embedding, device=device)
	else:
	return PyannoteAudioPretrainedSpeakerEmbedding(embedding, device=device)

It seems as a bit of a tight limit. For example, how about relaxing it as follows?

    if isinstance(embedding, str) and "speechbrain" in embedding.split("/"):
        return SpeechBrainPretrainedSpeakerEmbedding(embedding, device=device)
    else:
        return PyannoteAudioPretrainedSpeakerEmbedding(embedding, device=device)

I think when using the pyannote embeddings model, it is unlikely that the structure of directory contains "speechbrain".

[hbredin]

OK for making things a bit more flexible. I was not happy with the original solution anyway :-)

What about something even looser like if "speechbrain" in embedding?
This is still not perfect but that would probably do the trick...
I'd gladly merge a PR bringing this change.

[matakanobu]

What about something even looser like if "speechbrain" in embedding?

Much better.

This is still not perfect but that would probably do the trick...

Yes, this is still a workaround...

Sure. I will PR.

[hbredin]

I just merged the PR.