Add custom Whisper models for transcription

NoScribe comes with two preinstalled whisper models for transcription (precise/fast), both based on whisper V3 turbo. If you prefer to use the older whisper V2 model, or want to install a fine-tuned model specialized on a particular language, you can do that now (new in v0.6).

• Custom models are installed in the user config directory. In noScribe, click on 'Model > Add AI model...' to open this folder, or navigate there manually:
  • Windows: C:\Users\<username>\AppData\Local\noScribe\noScribe\whisper_models\ (Note that "AppData" is hidden in the Explorer by default. See here how to access it: https://www.howtogeek.com/885078/cant-find-appdata-folder-on-windows-11-how-to-get-it-back/)
  • Mac: ~/Library/Application Support/noscribe/whisper_models/
• Create a subfolder for each model. The name of the subfolder will show up as the name of the model in the noScribe user interface.
• Only models in the "Faster Whisper" format will work. They are sometimes also called "CTranslate2"-models, which refers to the technical foundation faster whisper is build on.
• Models in the original format from OpenAI (transformer based) can be converted.
• Download all the files from the model (usually: config.json, model.bin, preprocessor_config.json, tokenizer.json, vocabulary.json) into the folder you created. The new model should show up automatically in the noScribe UI.

List of models:

• The official faster-whisper repository (15 models): https://huggingface.co/Systran Here, you will find the original whsiper v2 model that was the standard up until noScribe version 0.5. The English-only models (suffix .en) are also interesting because they can give a significant speed bump.
• French fine-tuned version of whisper V3: https://huggingface.co/bofenghuang/whisper-large-v3-french/tree/main/ctranslate2 (if the downloaded files have a ctranslate2_ prefix, make sure to remove this, e.g., rename ctranslate2_vocabulary.json to vocabulary.json). You can also try the faster destilled version (untested): https://huggingface.co/brandenkmurray/faster-whisper-large-v3-french-distil-dec16
• German fine-tuned version of whisper V3 (my experience: only minor improvements in transcription quality, but less precision in pause-detection and the identification of overlapping speech): https://huggingface.co/aseifert/faster-whisper-large-v3-turbo-german/tree/main You will also need the file tokenizer.json from the original whisper V3 repository: https://huggingface.co/Systran/faster-whisper-large-v3/tree/main

The best place to find special models is huggingface. Search for "faster-whisper" and your language (or "whisper" only to find models that you can convert).

I am using this version of whisper V3 turbo as the "precise" model: https://huggingface.co/mobiuslabsgmbh/faster-whisper-large-v3-turbo

If you found a cool new model, please share it in this discussion: https://github.com/kaixxx/noScribe/discussions/132 (I might add it to the list above).