1h ago
Linux: local Whisper Large with Norwegian, Identify Speakers and system audio (parity with my Mac setup)
I run Superwhisper 1.0.2 on Linux (Omarchy/Arch, Intel Arc B390, Vulkan) next to my Mac. On the Mac my default mode is Whisper Large (formerly Ultra), fully local, language Norwegian, with Identify Speakers and system audio on. I use it for everything from dictation to calls, and my own tooling reads the speaker-labelled segments from meta.json. I can't reproduce that on Linux today:
1. Norwegian locally. The offline models on Linux (Cohere Transcribe, Parakeet v3) don't support Norwegian, so it falls back to English. Please ship Whisper Large as a local model on Linux. The bundled asr-worker already seems to include a Whisper backend: it loaded a Whisper ggml model on Vulkan in my test. Letting users point to their own ggml model, such as NB-Whisper from the National Library of Norway, would be a great bonus.
2. Identify Speakers for offline models, like on the Mac since v1.40.
3. System audio. PipeWire monitor sources are filtered out of superwhisper devices, so there's no way to capture the other side of a call. A per-mode "use system audio" toggle like the Mac's would solve it.
4. Same recordings format as the Mac (meta.json with segments and speaker ids), so tooling works across both machines.
Everything local. That's the reason I use Superwhisper.
Pending