0.9B local diarization model keeps speaker labels consistent across 17-minute gaps
Binary_orchid · reddit · 2026-11-05
A Reddit user stress-tested MOSS-Transcribe-Diarize 0.9B running locally on an M5 Ultra Mac Studio (fp32, greedy decoding, no hotwords) on 23m39s of a public National Capital Planning Commission meeting, verified against the agency's own verbatim transcript.
Results: 229.2 seconds total runtime, 309 segments, each with speaker label and timestamps. Six speakers who returned 17–19 minutes later were all re-tagged under the same labels without the model being told how many speakers existed, and a remote testifier's 2.5-minute statement stayed under one label end to end. A useful data point for anyone building meeting-transcription workflows on small local models.
More from Models
- Cognition launches SWE-2 Flash, a 276B-A12B coding model that beats all US open-weight rivals — carlesgelada · 2026-11-05
- Kimi K3.1 Pro (2.8T params, 104B active) lands on Ollama Cloud with zero data retention — ollama · 2026-11-05
- Linear CEO Karri Saarinen: time to retire the word "context" in AI — round · 2026-11-05
- Codex on the $20 plan runs dry after about 2.5-3 hours of nonstop work — tristanbob · 2026-11-05
- Morph launches Reflex: first decision model with cached pricing, 10-20x cheaper than Jev — GabGarrett · 2026-11-05
- Hamel Husain: Benchmarks are used like credit scores, so criticism is warranted — HamelHusain · 2026-11-05