Hi @Alykasym

The root cause was an ASR model that was always loaded to GPU (and left resident) even when a user supplied reference text

ASR model will not be loaded if ref_text provided.

Consider using a GPU card with >=4gb memory if you need ASR and TTS loaded together, otherwise the current PR's hardcoded modification will enforce reloading ASR model every time thus large overhead when inference.

Hi @SWivid , Thank you for quick response. I've tested on 8GB, 24GB and 48GB VRAM GPU. For some reason I was constantly getting skipped words especially in small sentences. But after explicitly adding unload ASR model, it never skipped a word.

the current PR's hardcoded modification will enforce reloading ASR model every time thus large overhead when inference.

I understand that, but keeping the ASR in memory leads to skipping words. Though no idea why exactly it happens.

And it doesn't skip words when I move ASR to CPU. As long as ASR not in the same memory with the F5 model, it doesn't skip words.

@Alykasym maybe it is better to provide us some reproducible buggy examples, i.e.:
the ref_text ref_audio gen_text and the used random_seed, while the output gen_audio is facing/not facing word skipping.

👀