Inquiry about Embedding Extraction

[PhilipAmadasun]

I have two questions about extracting embeddings

• First:
  from scipy.spatial.distance import cdist distance = cdist(embedding1, embedding2, metric="cosine")[0,0]
  Gives an error, I was following the tutorial here
  I have to do this instead:
  distance = cdist(np.expand_dims(embedding1,axis=0), np.expand_dims(embedding2, axis=0), metric="cosine")[0,0]
  To get results. And now I'm confused if I'm even supposed to do this.

• Second:
  Do you get the most accurate "Essence" of someone's voice via embedding by:
  - Extracting embedding from longer audio clips of the persons talking?
  - Gathering (let's say) 30 millisecond chunks of audio, extracting the embeddings from each chunk, then getting an average embedding from them?
  Which one of these is the way to go?

[fungus75]

First:
I guess they use nowadays speechbrain for identification of speakers: Link

Therefore the similarity-Identification might be done somehow like that/that.

I tried out the following:

similarity = torch.nn.CosineSimilarity(dim=-1, eps=1e-6)
sim = similarity(torch.from_numpy(embedding1), torch.from_numpy(embedding2))

But for me the question is, how big this "sim" must be to be identical.
According their code 0.25 is a good value? For me it is way too low.

[PhilipAmadasun]

@fungus75 The way I previously described still works when it's people whose voice embeddings I've already saved. I don't have a good threshold to pick put unknown voices. I'll try this way you've mentioned.

[PhilipAmadasun]

@fungus75 Yeah after some brief testing I'm not sure their way can work for me. Have you tried the way I mentioned previously and compared the two methods? I'm curious. I'm kinda working in briefs waves due to illness currently, would appreciate more input on this as I can't carry out this task currently. Maybe it it will also help you as well hopefully.