What's this about?
Chinese KOL dotey (Baoyu) documented a working recipe for fully-local speech-to-text: OpenMOSS's newly open-sourced MOSS-Transcribe-Diarize-0.9B produces speaker-labeled, precisely-timestamped transcripts of long audio in one pass, now wired into his BaoCut video tool. His earlier finding: Whisper (the usual default) has inaccurate timestamps and weak Chinese support, while Qwen3 ASR 0.6B + a forced-aligner model fixes both on a laptop.
What exactly is "diarization"?
Speaker diarization is figuring out *who* spoke *when* in a recording — splitting "60 minutes of audio" into "Alice: 0:00–0:42, Bob: 0:42–1:10…". Traditionally that required a separate model pipeline stacked on top of transcription. The interesting bit here is one small 0.9B model doing transcription + diarization + word-level timestamps in a single pass, small enough to run locally — the kind of boring-but-load-bearing capability that subtitle tools, meeting notes, and podcast apps are built on.
What's the community saying?
- Whisper's crown is slipping — for Chinese and mixed Chinese-English audio, the new small open models are simply better, and cheaper than cloud APIs.
- Known limitations shared openly — the MOSS model is slower than Qwen3 ASR, skips punctuation, and its custom-prompt/hotword features didn't seem to work; overlapping speakers still confuse diarization.
- Agent-skill packaging — the stack plugs into BaoCut as a skill so coding agents (Claude Code/Codex) can drive transcription end-to-end — a recurring dotey pattern.