Buzz Transcription: Speaker Count Control Lands
A single pull request, PR #1586, gives users manual control over speaker count in diarization, addressing a long-standing accuracy gap tracked in issues #379 and #1043.
Duration: PT2M10S
Episode overview
This episode is a short developer briefing from Buzz Transcription.
It explains recent repository work in plain language.
- Show: Buzz Transcription
- Published: 2026-08-19T13:16:22Z
- Audio duration: PT2M10S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good morning. It's August 19, 2026, and this is Buzz Transcription.
Today's focus is one change with real user-facing impact: control over how many speakers Buzz expects in a recording.
PR #1586 adds an Auto and two-to-eight speaker-count selector for the MSDD diarizer. Up to now, Buzz has relied entirely on automatic speaker estimation, which the PR notes can merge similar-sounding or quieter participants into one voice. That's a real accuracy problem for anyone transcribing group interviews or…
The fix passes a selected exact count through the worker layer into the vendored diarization engine. Worth noting: this doesn't apply universally. Sortformer, the other diarization backend in play, doesn't support a forced speaker count through its API, so the PR explicitly resets and disables the selector when…
The PR also documents the accuracy trade-off directly, which matters here: manually fixing a speaker count can help in known scenarios but may hurt if the guess is wrong. That's the kind of caveat that belongs in user-facing docs, and it's good to see it called out rather than buried.
On the engineering side, the change touches the speaker identification widget and its tests, along…
Nearby episodes from Buzz Transcription
- Cleaning Up Import Rules and Foundations
- Keeping Long Jobs Alive
- Speaker Identification Overhaul
- Speakers Get a Proper Data Model
- Weekly Recap - Cross-Platform Stability & Build Pipeline Cleanup
- Windows Playback Fix and CI Stability
- Speaker ID Gets a Fix, Recording Goes Network-Wide
- CI Pipeline Maintenance