Hi Maffers80, thanks for the detailed report, and sorry about that.
You were right: it wasn’t your AMD GPU. Version 2.7.0 had a bug and you found it. That's fixed in v2.7.1. Please update from the itch.io page and try again with that same file, it should work normally now.
Sorry for the hassle, and thanks again for buying it and taking the time to write this up. If anything still looks off after updating, let me know.
Thanks for the quick patch, I'll give it a try later. Is there any possibility of GPU acceleration working with AMD GPUs in the future? I have been using the program WhisperDesktop, which can use my AMD GPU (possibly using ZLUDA, but I'm not 100% sure of that) and runs about 10x faster than PTPro, but it is no longer being developed and it hallucinates random things too often. Also I see that PTPro uses v1 Language models, can we swap these out with newer models? Or do we need to stick with the ones provided?
Hi Maffers80,
GPU acceleration is done with Vulkan, which should work with AMD cards if the driver supports it. Here's some official info about Vulkan driver support for AMD
The models used in the app are simply a subset of the current ones. Largest bundled model is large-v1 as those were the ones that created the best outputs, and I don't want to increase even more the size of the app, the slider goes tiny→large-v1.
If you want to use a different model, you can do it by simply overwriting the model file itself (keep the same names though for the program to work). For example you can overwrite large-v1 with large-v2 and when the slider is at the most accurate one it will use large-v2 (it will load a model called large-v1 but the contents will be the large-v2 model).
That way you should be able to run any similar model that you want.
That works great, thanks so much. It is also working with my GPU (7900XTX) which has made it much faster.
One issue that I have noticed, both with PTPro and WhisperDesktop is that often a couple of lines of dialogue will get "stuck" and keep repeating. I understand that this might just be a weakness inherent to the whisper AI method, but wondered if you have tips for mitigating it?
Glad the GPU acceleration is working well on your 7900 XTX!
As you say, the stuck/repeating lines are basically coming from Whisper itself.
A few things that you could try on the app today:
- Try with a different model (smaller or larger). This sometimes fixes some issues.
- Set the spoken language explicitly in Settings instead of Auto Detection.
- Prefer cleaner dialogue audio when you can to avoid background noise.
There’s no perfect switch for it yet though, Whisper can still loop sometimes.
