MOSS Local adds on-device transcription with speaker labels. It streams short, closely timed segments but does not provide word-level timings, so long lines cannot be split as accurately.
Qwen Local and MOSS Local now run in separate worker processes, each with its own Python environment. They have incompatible transformers dependencies so they cannot be loaded into the same process. This may require qwen-asr to be reinstalled, but it makes things simpler and cleaner.
Runcrate is available as a translation gateway.
What's Changed
- Add MOSS Local transcription provider and local runtime setup wizard by @machinewrapped in #547
- Support anthropic 2.x and mistralai 3.x SDKs by @machinewrapped in #545
- Run Qwen Local in a worker process by @machinewrapped in #544
- Flag repeated response labels in translated subtitles by @machinewrapped in #546
- perf(transcription): 1.47x faster IsSilent peak scan (silence gate test) by @dexhunter in #548
- perf(transcription): 12.9x faster AlignWords on 10 min transcript chunks by @dexhunter in #549
- perf(transcription): 11x faster _best_split_index for long utterances by @dexhunter in #552
- feat(providers): add Runcrate gateway by @heydryft in #540
- Group transcription providers by cloud and local by @machinewrapped in #550
- Add visibility controls for gateways and cloud providers by @machinewrapped in #555
New Contributors
- @dexhunter made their first contribution in #548
- @heydryft made their first contribution in #540
Full Changelog: v1.7.4...v1.8.0