Skip to content
MinusPod
ThemeNord

Same themes as the MinusPod app

Default

Catppuccin

Dracula

Nord

Gruvbox

Solarized

Rose Pine

Kanagawa

Ayu

Other

Episode Processing Workflows

Every way MinusPod can process an episode, in pictures. For the behaviour behind each stage, see How It Works.

The job

Section titled “The job”
A published episode with pre-roll, mid-roll and post-roll ads, and the shorter file MinusPod serves in its place A published episode with pre-roll, mid-roll and post-roll ads, and the shorter file MinusPod serves in its place

The listener subscribes to a MinusPod address once. Everything below happens behind that address.

How work arrives

Section titled “How work arrives”
Scheduled poll, publisher announcement and listener playback all feed one queue that processes a single episode at a time Scheduled poll, publisher announcement and listener playback all feed one queue that processes a single episode at a time

The Whisper pool

Section titled “The Whisper pool”
A waiting queue of episodes feeding a Whisper pool that transcribes four at once, which then hand off to the pipeline A waiting queue of episodes feeding a Whisper pool that transcribes four at once, which then hand off to the pipeline

The four concurrent episodes in the diagram are an example, not the setting. Concurrency is configurable: WHISPER_POOL_MAX_EPISODES accepts 1 to 16 and defaults to 1, and out-of-range values are clamped rather than rejected. A separate cap, WHISPER_POOL_MAX_REQUESTS (1 to 64, default 4), limits how many transcription requests are in flight at once.

By default one episode runs at a time. Turn on the Whisper pool and raise its per-episode limit to transcribe several at once; the rest wait in the queue and start as slots free. Each running episode still flows through the same pipeline once its transcript is ready. The pool is off by default.

Whatever the backend, only the background leader worker runs episodes. A reprocess or play request that lands on another gunicorn worker is queued and the leader’s dispatcher picks it up on its next pass, within 30 s. One process holds every per-run state and every local transcription model, so two workers can never load a model each on the same GPU.

The pool only activates on the remote Whisper API backend. Turning it on while transcription runs locally leaves it inactive, because the local model runs in-process and several episodes would contend for the same GPU. An inactive pool reports a per-episode limit of 1 whatever the setting says, so the queue keeps running one episode at a time.

The standard pipeline

Section titled “The standard pipeline”
Eleven pipeline stages from download through publish, with the second download and audio analysis running in parallel Eleven pipeline stages from download through publish, with the second download and audio analysis running in parallel

Amber badges mark the stages that make an LLM call. Transcription is billed too if you point it at a hosted Whisper API; on a local GPU it costs only time.

Long episodes transcribe in chunks sized to fit GPU memory. Cutting and normalizing a chunk is its own ffmpeg pass, and that pass runs ahead of the GPU: while one chunk transcribes, the next ones are already being prepared. The verify pass re-transcribes the cut file the same way.

The YouTube pipeline

Section titled “The YouTube pipeline”
Six stages from top to bottom: discover, serve, wait, cut, transcript and serve the cut. Beside the wait stage, marks found are queued and no marks leads to an LLM fallback or watch window. A background backfill fills descriptions and times Six stages from top to bottom: discover, serve, wait, cut, transcript and serve the cut. Beside the wait stage, marks found are queued and no marks leads to an LLM fallback or watch window. A background backfill fills descriptions and times

A YouTube feed has no upstream RSS, so MinusPod builds the episode list itself and waits for community SponsorBlock marks before it cuts. Detail for every step is in YouTube Feeds.

  • Discovery. On the normal feed refresh schedule MinusPod lists the playlist, channel or single video. It uses the YouTube Data API when a key is set and yt-dlp otherwise. Private and deleted entries are hidden.
  • Serving. The served feed lists every discovered episode with its upload date. Nothing is processed at discovery: the first play of an unprocessed episode starts processing, and the podcast app is told to retry until the cut is ready.
  • SponsorBlock gate. An episode waits the processing delay, then MinusPod probes SponsorBlock a set number of times at a fixed interval. The first probe that finds marks queues the episode. If none appear after the last probe, the LLM pipeline runs as a fallback. With the fallback off, the episode enters a watch window and is checked at the watch interval until the watch ceiling.
  • Cutting. Marks are cut as given, with no LLM call by default. Learned patterns (audio fingerprints and text patterns) always run, and MinusPod learns new patterns from the marks it cuts. Pass 1 detection, a per-feed switch, adds the LLM on top and cuts both sets of segments. If the marks change during the watch window, the episode is recut from the retained original.
  • Transcript. When captions exist and the captions setting is on, they become the transcript and Whisper is skipped. Creator chapters are reused, so the chapter generation call is skipped too.
  • Metadata backfill. A background pass fills missing descriptions and exact publish times from each video’s metadata, for up to 10 episodes per refresh pass, without downloading.

Global settings live in Settings > YouTube. The delay, probe, fallback, captions, private-video hiding, transcription fallback and Shorts settings can be overridden per feed, while the watch interval and ceiling are global only. See Configuration for the full table.

Five kinds of evidence, five outcomes

Section titled “Five kinds of evidence, five outcomes”
Acoustic match, known script, copy comparison and audio signals are free; the model call is paid. All five feed one decision gate that cuts, beeps, keeps, holds or rejects Acoustic match, known script, copy comparison and audio signals are free; the model call is paid. All five feed one decision gate that cuts, beeps, keeps, holds or rejects

Four of the five cost nothing. Fingerprints, learned scripts and the two-copy comparison are gathered before the model reads the transcript, and go into its prompt as hints. Audio signals are measured then too, but only promote or extend a cut after detection runs. Whatever a source proposes still has to clear the gate.

System One scores a window as one request with one question per line; see the diagram below and System One.

A transcript split into overlapping windows, where the overlap is scored twice. One window becomes one request: a shared state, one question per line split into batches that carry the same state, and a prefill per batch. A max prompt tokens budget splits oversized windows. Yes or no answers feed the decision gate, which cuts, holds or keeps A transcript split into overlapping windows, where the overlap is scored twice. One window becomes one request: a shared state, one question per line split into batches that carry the same state, and a prefill per batch. A max prompt tokens budget splits oversized windows. Yes or no answers feed the decision gate, which cuts, holds or keeps

Processing modes

Section titled “Processing modes”

One mode per feed. Changing it does not change the published address. A single episode can also be set to pass-through on its own, which overrides the feed mode for that episode only: it skips transcription, detection, and editing. The audio may still be transcoded for serving, so pass-through means no ad removal rather than byte-identical audio. Set it from the episode page’s Reprocess menu, or select episodes on the feed page and use the bulk pass-through action.

Standard, keep content only, cue-only, skip ad detection and pass-through, each with the shape of audio it produces Standard, keep content only, cue-only, skip ad detection and pass-through, each with the shape of audio it produces
Stage Standard Keep content Cue-only Skip detection Pass-through
Transcribe yes yes optional yes no
Second download per feed per feed per feed no no
Audio analysis yes yes yes no no
Model reads transcript yes inverted no no no
Learned patterns match, learn match only match, learn no no
Verify pass per feed per feed no no no
Audio edited yes yes yes no no
Transcript, chapters yes yes optional yes no

Keep content safety net. Marked content has to cover at least 55% of the runtime, no single cut may exceed 7 minutes or a quarter of the episode, and every transcript window has to come back labelled. Miss any of those and the episode falls back to standard detection on its own.

One episode under each of the five modes, then segment categories and the remove, beep, keep and mark actions. Portrait cut and the file itself in Videos.

Re-running an episode

Section titled “Re-running an episode”
Reprocess, full analysis, re-detect ads and recut audio, showing which pipeline stages each one runs, reuses or skips Reprocess, full analysis, re-detect ads and recut audio, showing which pipeline stages each one runs, reuses or skips

Reprocess and full analysis reuse the saved transcript when there is one, and transcribe again when there is not. Re-detect ads always needs one, so it skips episodes that have none.

Learning loop

Section titled “Learning loop”
Detected ads become stored patterns that match for free on later episodes, widening from one show to a network to global scope Detected ads become stored patterns that match for free on later episodes, widening from one show to a network to global scope

Every pattern carries its own confirmation and false-positive counts. One you created or confirmed always cuts what it matches; one the system learned on its own still answers to the category actions set for the feed.

Community sync is opt-in in both directions. Nothing local is published unless you submit it.

When something breaks

Section titled “When something breaks”
A failed model call publishes partial cuts, an offline endpoint parks the episode, and an uncertain cut is held rather than applied A failed model call publishes partial cuts, an offline endpoint parks the episode, and an uncertain cut is held rather than applied

Nothing fails silently. A run with no free evidence to fall back on, or one that died on a bad API key, fails outright and waits for you on the episode page. The offline queue is off until you turn it on; without it, an unreachable endpoint retries and then gives up.

Diagrams are generated, not hand-drawn. Edit scripts/generate_workflow_diagrams.py and re-run it to rebuild docs/images/wf-*.svg. Colors come from the index.css tokens, so they track the app’s own theme.