Feature Request: Auto convert custom whisper models to Faster whisper format when downloading

[Cherkah]

Description

Buzz currently only supports standard Faster-Whisper models (tiny/small/medium/large + .eng variants), limiting language-specific accuracy for non-English audio.

I'd like to request the ability to:

1. Download/import custom Whisper models directly from Hugging Face repositories
2. Use fine-tuned models like yuriyvnv/whisper-small-mixed-cv-nl for superior Dutch transcription (WER ~13% vs ~18% multilingual small) .

Use case

• Located in Netherlands (Roosendaal), need accurate Dutch speech-to-text
• Recent HF models (2025) outperform multilingual baselines significantly on accents/bruit
• Current workflow: CLI faster-whisper → SRT import in Buzz (inefficient)

Proposed implementation

• Settings > Models > "Download from Hugging Face" field (repo ID)
• Auto-conversion to Faster-Whisper format with CUDA/NVIDIA support
• Model cache in ~/.local/share/buzz/models/

Current limitations

• No custom model support found in docs/FAQ
• No related feature requests in Issues/Discussions (searched 2023-2026)

Screenshots (if applicable)
[Attach Buzz model selection screen]

Environment

• OS: Fedora 43 KDE Plasma 6.5 + NVIDIA (RPM Fusion)
• Buzz: [version from flatpak list]
• GPU acceleration working correctly with standard models

[raivisdejus]

Faster whisper supports custom models. If you select custom model size a text field will appear for you to enter the huggingface model ID. But the model has to already be in the Ctranslate2 format on huggingface. Easiest way to find such models is to search for "Automatic Speech Recognition" task + whisper + ct2 - https://huggingface.co/models?pipeline_tag=automatic-speech-recognition&sort=trending&search=whisper+ct2

Some useful custom models I have collected here #866

But I really like the idea about auto converting regular models to the faster whisper format and I agree that the documentation could be improved.

[Cherkah]

Follow-up: Additional Usability Improvements

Thanks for the quick response and model support clarification! While testing realst1112/whisperd-nl-ct2 and Systran/faster-whisper-small, I identified these usability issues:

1. Slow Startup (5-6s delay)

• Issue: Buzz takes 5-6 seconds to launch on Fedora 43 KDE Plasma 6.5 + NVIDIA
• Expected: Instant launch (<1s) like other Flatpak apps

2. Missing Reordering Arrows

• Issue: No up/down arrows to reorder completed transcriptions in the list
• Expected: Drag&drop or arrow buttons for priority editing/export
• Use case: Better files management

3. File Picker Path Inconsistency

• Issue: Open File dialog defaults to ~ instead of following "Output folder" from Watch Folder settings
• Expected: File picker remembers last used path / syncs with output/watch folders
• Workflow: Select input → set output → repeat without path hunting

4. Custom Model Naming

• Issue: Downloaded custom models show generic "custom" name in dropdown
• Expected: Option to rename (ex: "NL-whisperd" vs "NL-Systran-small")
• Impact: Hard to distinguish between multiple custom models tested

5. Preferences Menu Location

• Current: Help → Preferences (non-intuitive)
• Expected: Top-level Preferences menu item next to File/Edit
• Standard: Matches KDE/GTK conventions

Environment: Fedora 43 KDE Plasma 6.5, NVIDIA GPU (RPM Fusion), Flatpak Buzz latest

These are small but high-impact polish items. Happy to test PRs! Thanks for the great app.