improved model

[mrdrprofuroboros]

Just in case anyone is still interested, here's an improved embedder. Thanks for your work, @davidbrowne17
https://github.com/mrdrprofuroboros/unmute-voice/tree/main

[davidbrowne17]

Cool! I uploaded a new and better version, here are the training results

• MSE Loss: 0.050422
• Cosine Loss: 0.051857
• Distribution Loss: 0.001720

Let me know what you think

[mrdrprofuroboros]

Neat, that’s quite an improvement!
Is that still adapter only? How did you train it?
Does the resulting embeddings any meaningfully better?

[davidbrowne17]

Yep its the adapter only, Ill release training code at some point but im waiting for some sort of feedback from kyutai as I dont want to get them into legal trouble since they and I are in the EU and laws are stricter (Thats the reason they didnt release the model in the first place). From my experience the voices are very nice and it can capture certain "popular" voices very well. I have many versions of it so id love to hear more feedback on its weak points so I can do more testing