[dben]

An optional second container that transcribes episodes, so the private feeds carry captions. pod-block works as before without it: the quick start, the main image and kubectl apply -k deployment/ are unchanged.

The transcription service (whisper/)

• faster-whisper on the CPU (int8), behind one streaming HTTP endpoint. It works on one episode at a time and stops when the client hangs up.
• Transcribes in batches of 8 by default (WHISPER_BATCH): about twice as quick on a CPU, and it punctuates better.
• Optional speaker turns (WHISPER_SPEAKERS=true) via sherpa-onnx: pyannote segmentation and a WeSpeaker voice model. No PyTorch or Hugging Face token; the models (about 35 MB) are fetched on first start. POST /speakers re-runs that step alone with other clustering options.
• Every package is pinned. faster-whisper 1.2.1 breaks with PyAV 19, so av stays at 18.
• CI builds and publishes ghcr.io/<repo>-whisper beside the app's image.

In pod-block

• PODBLOCK_WHISPER_URL turns it on. Each episode cut without a timed transcript from its feed gets a captions job, which sends the uncut download to the service.
• The words come back with times and are split into captions of at most two 42-character lines. They're saved as WebVTT (rel="captions") among the episode's transcripts, so the existing code shifts them to fit the cut and serves them; a re-cut doesn't need a new transcript.
• The job names its episode in its payload rather than holding it, so the episode stays ready and in the feeds while it runs. Captions jobs get a worker thread of their own, so transcriptions don't hold up downloads.
• Speaker turns, when the service sends them, are saved beside the captions. They stay out of the captions and the cuts for now. podblock speakers EPISODE_ID compares them with the episode's repeats: each voice's talk time, its share inside the repeats, voices heard mostly inside the repeats, and lone voices. --threshold and --count re-run only the speaker step, for tuning.

Deployment

• deployment/whisper/ is a separate kustomization with its own labels, so pod-block's Service never selects its pods. The example uses large-v3-turbo with 6 threads.
• deployment/README.md has a new "Optional: captions" section covering adding and removing it.

Measured on 4 cores (2.8 GHz Xeon), real episodes

	Share of episode length
small, one at a time	0.38
small, batches of 8	0.20
large-v3-turbo, batches of 8	0.43
speakers	0.17

On three real shows, two read their own ads, so speaker turns had nothing to add. On a network show, threshold 0.65 separated the ad voices from the hosts (two voices 67% and 88% inside the repeats), while 0.8 merged most of them into the hosts. The default threshold needs more tuning before anything is cut by it.

Testing

• The full suite passes (201 tests), and ruff is clean.
• The service's real HTTP handler is tested with stand-in models, end to end through the captions job.
• I also ran the real models end to end, and built and ran the image with a read-only root file system as uid 1000.

---

Generated by Claude Code