Dispatch 223
Week of May 11 to May 17 ↗

Chunking long audio so the GPU doesn't cry

Monday started with a real problem: Parakeet would OOM on long recordings. Ten minute chunks fixed it. I split the audio into 10 minute pieces to avoid CUDA OOM, a one file change in the parakeet service tha…

Commits
2
Systems
1
Read
2min
MAY 11, 2026Monday345 WORDS · 2 MIN READ · 2 COMMITSuc-meeting-ops2
Day signal

Drawn from this entry's own record: its date, its length and the work around it. No stock art.

Monday started with a real problem: Parakeet would OOM on long recordings. Ten minute chunks fixed it. I split the audio into 10 minute pieces to avoid CUDA OOM, a one file change in the parakeet service that adds a chunker and lets the model process anything without blowing up memory. The number that mattered was the crash count dropping to zero.

Before this, the transcription pipeline had a hard cap you hit silently. Record a 90 minute meeting and watch it fail. The chunking is simple and effective. Feed it a long file, it breaks it apart, processes each piece, then stitches the results back together. No drama.

The rest of the day was about making Parakeet the default for everyone, not just the people who knew how to flip a config flag. I wired it into the provider registry, added a model picker in the frontend panel, and added a thinking toggle for the LLM side. Six files changed. The transcription options panel now shows what you're actually using instead of a hardcoded provider name.

I wired it into the provider registry, added a model picker in the frontend panel, and added a thinking toggle for the LLM side.

Diarization got a four hour cap. Anything longer than that needs manual handling. I put it in the transcription options service so the validation lives next to the rest of the constraints rather than hiding somewhere in the provider code.

The deploy compose file for bigboy got updated to make sure the parakeet defaults land in production the right way. No new services, just the right defaults at the right layer.

Also today: the provider registry got cleaned up so adding the next model won't require touching three different files. The LLM provider implementation and the transcription options module are now aligned on the same shape, which makes the next provider addition just a registry entry and a config line.

The day added up to a pipeline that handles long audio without crashing and a UI that actually tells you what model you're running. That's the kind of improvement that doesn't show up in a demo but keeps the system from embarrassing us at 4pm on a Friday.

From the recordWhere the work went2 commits across 1 repository
  1. uc-meeting-ops2