While OpenAI Whisper is great, it has its own shortcomings, such as hallucinations, and some of the languages it supports are not great. Which is why i suggest adding more local model options, something like Parakeet STT by Nvidia which focuses more on European based languages, or Vibevoice by Microsoft, which has more accuracy. These alternatives may be better then OpenAI’s Whisper at different languages and scenarios.
Alternatively, add some sort of (maybe experimental?) option to import models from hugging face. This would allow a user to use a specific model trained for a specific language that would be a lot more accurate then general STT models.
Are they compatible with whisper.cpp? We cannot justify to integrate more runtimes due to the effort, limited human resources, increased build time and maintenance, increased install size, and priority (niche feature). And we do not have a generic UI add-on system. So, unlikely even if contributed by an outsider.
Unfortunately they aren’t compatible with whisper.cpp, but what about importing fintuned whisper models? these should work with whisper.cpp since they are built on top of whisper and still have the same underlying structure. You can add a disclaimer that because these are not the official models they may not work and cause weird issues.
Have you see the documentation that explains that you can download any compatible GGML model from Hugginface and configure Shotcut to use it?
Just click on the “Configure” button and specify the file you want to use. If you find some models that work good for certain situations, I’m sure people would be interested to know about your experience. If they are generally useful, I can add them to the list of models that Shotcut can download for a future release.
Yeah that third reply has a point about GGML models already working, but there’s a gap between “technically possible” and “actually low-friction for most users.” Converting a fine-tuned Whisper from Hugging Face to GGML format and then configuring it isn’t exactly a one-click job. That documentation link might be the answer here, but if it doesn’t walk people through getting a model *into* GGML format first, a lot of users will bounce off it.
I completely forgot that you can use GGML models via just selecting them from my system with the file picker. I think it should be a bit more clear that you can select custom GGML models (Add some sort of text explaining that you can select a custom GGML model) and when you select a custom model it gets saved into the list of speech to text models that you can use. So that way if you have multiple ones that you use for multiple different things, you can just quickly select between them.
There is already a clearly labeled field in the configuration section. I guess what you are requesting is that the configuration section is always shown instead of hidden by a toggle. By that argument, why not show every button and option of everything on screen at the same time making it look like the cockpit of a commercial jet?