Repository navigation
- Parakeet Ultra on any x86-64 CPU: the AVX2 requirement from 2.5.0 is gone. (#145)
- Smaller downloads: the Ultra model goes from 486 MB to 416 MB and keeps its accuracy, 1.81% WER on LibriSpeech test-clean against 1.80% for Moondream's own runtime. The AppImage is about 15 MiB smaller, the Flatpak about 8 MiB and the installed Fedora package 38 MiB. (#145)
- One speech engine: Ultra now runs on whisper.cpp like the other models, so Dabri no longer bundles transcribe.cpp and OpenBLAS. (#145)
- Fix:
dabri stopprints the transcript right away when typing and the clipboard both fail, instead of waiting 45 seconds. (#146) - Config:
whisper_modelin the config is nowmodel. Dabri reads the old key and writes the new one the next time it saves the config. (#143) - Packaging: one release workflow builds and publishes every package, RC tags build them without publishing, and the RPM spec follows the Fedora packaging guidelines. (#142, #144)
If you use Parakeet Ultra, Dabri downloads the new model the next time it loads it. You can delete the old parakeet-ultra-Q4_K_M.gguf from the models folder.
The converted model and the conversion script: https://huggingface.co/AshBuk/dabri-models
What's Changed
- release: one pipeline for stable and rc tags, docs sync by @AshBuk in #142
- config: rename whisper_model to model by @AshBuk in #143
- packaging: follow Fedora guidelines in the RPM spec by @AshBuk in #144
- stt: run Parakeet Ultra through libparakeet and drop transcribe.cpp by @AshBuk in #145
- services: return the transcript to dabri stop when output fails by @AshBuk in #146
Full Changelog: v2.5.0...v2.6.0