Maksym Sutkovenko

Yeap problem reproduced if virtual env is used.

Good question.

Could you provide maybe some insights what layers should be frozen in decoder to fine tune the model without reseting decoder weights to zeroes?

I have been experience the same problem.

Okay, it looks like I figured out what's problem.

On my Raspberry Pi Zero 2 W, by default I have two devices: ``` card 0: vc4hdmi card 1: wm8960soundcard ``` You can check this using `aplay -l`. Then I...

Maybe we need some code that would either disabling vc4hdmi or setting proper card by searching in system wm8960soundcard.

What about using SPI0 instead of SPI1?

@ex3ndr Flow Matching paper recommends to set it sufficiently small. Guys from MatchaTTS(which also uses Flow Matching) trained their model with 1e-5 value.

Experience the same problem.