Feature Request: Download and select language models individually in-app or during installation

[candideu]

Hi there! I wanted to relay a feature suggestion shared in David Mbugua's review of the app: giving people the option to select which size of language model they want, and which language they will use.

This could be an installation step, where people select them, or the models could be downloaded from inside the app after the app is installed.

Here is an example from Audapolis:

Another example from GPT4ALL:

[kaixxx]

Whisper is actually multilingual. There are no models only for French or German.
Exception: Some Englisch-only models are available.

Regarding the selection of models:

• noScribe comes with two models: The quality setting "precise" uses the large model, "fast" uses the small model of whisper.cpp.
• If you really want to change these models, you can do this too: Download the models you want from huggingface. Then edit the “config.yml” as described under “Advanced Options” in the noScribe-Readme so that the quality options “precise” and “fast” point to the new models.
• Problem is that there are several versions of these models out there, some of which are not compatible with noScribe. This is also why I decided to include the models in my download.

[PedroMD]

+1 here. I'm on a Mac, and after inspecting the contents of the .app I found that the latest build does not include v3 of whisper. I might try to download v3 from HF and update the config.yml, but the ability to select a few models from the app would be extremely useful.

[kaixxx]

Hi Pedro, only the large whisper model has been updated to version 3. From what I've heard, the results are mixed. Some people even reported the performance has degraded compared to v2 (more hallucination). In my own testing, I didn't see much difference between the two models. So I didn't bother to upgrade.
I'm not sure if simply replacing the model will work, but you can give it a try. The link above is no longer valid. I've since switched from whisper.cpp to faster-whisper as the underlying library, so you must use models in their format: https://github.com/SYSTRAN/faster-whisper. Please report back if you had any success.
I might make the selection of different models from the UI a little easier in the future (allowing for extra options besides 'fast' and 'precise'), we will see.