Summarizing a 1-hour recording with Ollama: qwen3:8b took about 15 minutes. How do you handle long transcripts?
retained text
Disclosure: I'm posting this on behalf of the Nomily team. They make a recording app and sell the V05E recording card, so there's a commercial interest here. The team has been testing a local pipeline: the card records, the audio goes to the phone over Bluetooth, faster-whisper on a computer on the LAN does the transcription, and Ollama on the same computer does the summary. With this setup, audio and transcripts only go to machines on your own LAN. Test setup: \- Computer: MacBook Pro (Intel, 32 GB RAM) \- Transcription: faster-whisper \[model name\] \- Summary: qwen3:8b \- Recording: 1 hour, \[language / scenario\] Results: \- Transcription: about 20 minutes \- Summary: about 15 minutes Things that tripped us up: \- The phone can't use localhost. You have to enter the computer's LAN IP. \- Ollama needs OLLAMA\_HOST=0.0.0.0:11434 so the phone can reach it. Ollama has no authentication of its own, so any device on the same network can call it. Only do this on a network you trust. \- This local faster-whisper setup has no speaker diarization, which hurts summaries of multi-person meetings. Two questions: 1. For a 1-hour transcript, do you summarize in chunks and merge, or just raise num\_ctx? 2. For meeting summaries with models around 8B, is there anything more reliable than qwen3:8b?
document fetched 2026-11-20 00:41 · canonical https://reddit.com/r/ollama/comments/1y05yoy/summarizing_a_1hour_recording_with_ollama_qwen38b