Follow-up to #33. docs/research/audio-length-ceiling-2026-10-01.md records what was measured; this is what it did not:
No ceiling was found. 90 minutes (174 MB WAV, one request, 440 s) passed on iguana/kb-whisper; the 90-minute cap in TAPIR_UPLOAD_MAX_MINUTES is the longest length tried, not a limit. Raising it is a measurement, not an edit.
Concurrency. One job at a time is deliberate; two simultaneous 90-minute jobs on iguana were never tried.
berget/kb-whisper beyond 22 min / 42 MB — enough to know the fallback is not capped at meeting length for 22 minutes, nothing more.
Real speech. The fixture is single-voice macOS TTS looped from an 82-second script. Accuracy, crosstalk, and real density (chars/minute) are untested. The 90,000-char chat budget assumes the fixture's 828 chars/min.
The Metal OOM (infra#273) was deliberately avoided. A large resident ollama model during a long transcription is still the known way to wedge iguana permanently — and the default chat model for uploads (iguana/gemma4-26b) is exactly that kind of model on the same host.
A real Teams export: container, codec and channel layout of an actual voice-only .mp4 was not tested (only a synthetic AAC .m4a).
Suggested: run one real (non-committed!) recording end to end once #31's custody decision exists, and re-measure density.
Follow-up to #33. `docs/research/audio-length-ceiling-2026-10-01.md` records what was measured; this is what it did **not**:
- **No ceiling was found.** 90 minutes (174 MB WAV, one request, 440 s) passed on `iguana/kb-whisper`; the 90-minute cap in `TAPIR_UPLOAD_MAX_MINUTES` is *the longest length tried*, not a limit. Raising it is a measurement, not an edit.
- **Concurrency.** One job at a time is deliberate; two simultaneous 90-minute jobs on iguana were never tried.
- **`berget/kb-whisper` beyond 22 min / 42 MB** — enough to know the fallback is not capped at meeting length for 22 minutes, nothing more.
- **Real speech.** The fixture is single-voice macOS TTS looped from an 82-second script. Accuracy, crosstalk, and real density (chars/minute) are untested. The 90,000-char chat budget assumes the fixture's 828 chars/min.
- **The Metal OOM** (infra#273) was deliberately avoided. A large resident ollama model during a long transcription is still the known way to wedge iguana permanently — and the default *chat* model for uploads (`iguana/gemma4-26b`) is exactly that kind of model on the same host.
- **A real Teams export**: container, codec and channel layout of an actual voice-only `.mp4` was not tested (only a synthetic AAC `.m4a`).
Suggested: run one real (non-committed!) recording end to end once #31's custody decision exists, and re-measure density.
Refs: #33, #29, infra#273.
---
_Created via git-mcp on behalf of @mathias_
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Follow-up to #33.
docs/research/audio-length-ceiling-2026-10-01.mdrecords what was measured; this is what it did not:iguana/kb-whisper; the 90-minute cap inTAPIR_UPLOAD_MAX_MINUTESis the longest length tried, not a limit. Raising it is a measurement, not an edit.berget/kb-whisperbeyond 22 min / 42 MB — enough to know the fallback is not capped at meeting length for 22 minutes, nothing more.iguana/gemma4-26b) is exactly that kind of model on the same host..mp4was not tested (only a synthetic AAC.m4a).Suggested: run one real (non-committed!) recording end to end once #31's custody decision exists, and re-measure density.
Refs: #33, #29, infra#273.
Created via git-mcp on behalf of @mathias