Glossary
Every term the app uses, in plain words, with a link to the part of the docs that covers it. If you hit a word in the UI that isn’t here, open an issue.
Ad chapter - A chapter MinusPod adds for a segment it left in the audio (kept by category action, or waiting for review) so a chapter-aware player can skip it. Podcasting 2.0 > Ad chapters
Ad Review - The Patterns page tab that lists every detection across all feeds so you can confirm or reject them in one place, with a play button for the flagged audio. Web Interface > Ad Review tab
Ad Reviewer - An optional second LLM that double-checks each planned cut before it happens and can confirm, adjust, reject, or resurrect a detection. Off by default. Configuration > Ad Reviewer
Addressing mode - How the LLM names an ad’s position: timestamps (it invents start and end times) or segment_ids (it names numbered transcript lines that MinusPod maps to exact times). The Ad addressing mode setting is experimental and random draws one per run. Each LLM-detected marker records the mode it was detected under (addressing_mode) and the model’s raw bounds (llm_start, llm_end), and the stats API compares reviewer verdicts, correction rates and snap distance per mode. Configuration > Ad Addressing Mode
Archive mode - A per-feed retention setting that keeps every processed episode indefinitely, ignoring the global retention window and the “Clear all processed audio” action. Set it on shows you never want swept, such as ones that have stopped publishing. Configuration > Per-feed retention
Audio analysis - A pre-detection pass over the audio itself (volume shifts, transitions, silence) whose signals feed the detector and validator. How It Works > Audio Analysis
Audio cue - A short, repeated sound a show plays around its ad breaks (a sting, a jingle). MinusPod can learn one as a template and use matches as hard evidence for ad boundaries. Audio Cue Detection
Authenticated feeds - An optional key appended to feed URLs so only clients holding it can pull your processed feeds. Security > Authenticated feeds
Cache hit rate - Share of input tokens served from the prompt cache, computed as cache reads over input tokens. Shown on the Stats page per model, per run and per contributing call, with the raw read, write and input counts on hover. Web Interface > LLM cost ledger
Captions as transcript - Using the video’s own caption track as the episode transcript instead of running Whisper. Creator captions win over auto-generated ones, in the feed’s language, and a global setting plus a per-feed override turn it on or off. YouTube Feeds > Transcripts and chapters
Chapters - Chapter markers MinusPod preserves or generates for the processed audio, served as Podcasting 2.0 JSON and embedded in the MP3 as ID3 frames. How It Works > Chapter Generation
Chapter regeneration - Refreshing an episode’s chapters on demand. Each attempt gets its own processing-history entry and usage record, plus an episode log when log storage is enabled. How It Works > Chapter Generation
Cleanup - The Patterns page tab for approving, rejecting, and undoing suggested edits to learned patterns. These decisions change the pattern library, not audio already published. Pattern Cleanup
Community patterns - Opt-in sharing of learned ad patterns with other MinusPod users, and pulling theirs in return. Configuration > Community Patterns
Confidence - The detector’s certainty (0-100%) that a flagged region is an ad. Only detections at or above the cut threshold are removed; the rest stay in the audio for review. Configuration > Ad Detection Settings
Configuration Import / Export - Exporting and restoring runtime global settings and feed configuration. It preserves episodes, history, metrics, and media. Configuration Import / Export
Contaminated pattern - A learned ad pattern that contains show content or other material that should not be matched as an ad. Pattern Cleanup can flag it for review; nothing changes until a suggestion is approved. Pattern Cleanup
Cookies file - An optional Netscape-format cookies.txt from a throwaway account that yt-dlp uses for videos that need a signed-in session. Stored on the data volume with mode 0600 and never logged. YouTube Feeds > Cookies
Correction - Your verdict on a detection: “Confirm” or “Not ad”. Corrections train future detection for that feed and can trigger a recut. Web Interface > Ad Review tab
Corroboration - Independent evidence from another detection stage, an overlapping marker, or a matched audio cue that backs an otherwise-uncertain detection, letting it cut instead of hold. Used by the cross-fetch differential stage to decide which candidates cut and which hold. How It Works > Cross-Fetch Differential
Cover art badge - The small mark MinusPod adds to a feed’s artwork so you can tell the processed feed apart from the original in a podcast app. Configuration > Cover art badge
Cross-fetch differential - Downloading an episode twice and comparing the copies; audio that differs between fetches was inserted by an ad server, not part of the show. Runs automatically on feeds that look DAI-served. How It Works > Cross-Fetch Differential
Cue match - A place in an episode where a learned audio cue template was found, shown with its score and how detection used it. Audio Cues > Cue matches on an episode
Cue-only - The experimental per-feed Processing mode that cuts from cue pairs and previously learned ad patterns, with no LLM reading the transcript. Requires an enabled ad-break-start and an enabled ad-break-end template on the feed. Audio Cue Detection > Cue-only preset
Cue pair - Two bracketing template cues, an ad-break-start and an ad-break-end, that propose an ad span without reading the transcript. Each cue is included in the cut unless its removal option is off. Cutting from cue pairs is experimental and off by default. Audio Cue Detection > Ad cutting
DAI (dynamic ad insertion) - Ads spliced into the audio by the publisher’s server at download time, so every download can carry a different ad load. This is why the same episode can be 46 minutes one fetch and 55 the next. How It Works > Cross-Fetch Differential
Defined pattern - A user-created or synced community ad pattern. Its match always cuts the matched segment, overriding whatever action the segment’s category resolves to. An auto-learned pattern is not defined and still respects segment actions. How It Works > Pattern Learning
Deno - The JavaScript runtime yt-dlp needs to solve YouTube’s player challenges. Both images install a pinned, checksum-verified copy. YouTube Feeds > Troubleshooting
Detections Not Cut - The episode page section listing regions the detector flagged but left in the audio: rejected by validation, below the confidence threshold, or vetoed by the reviewer. How It Works > Post-Detection Validation
Deferred - An episode parked because the LLM or transcription endpoint was unreachable. It retries automatically when the endpoint comes back instead of burning through its retry budget. Configuration > Offline Queue
Duration filter - A per-feed minimum or maximum RSS duration that skips automatic and just-in-time processing outside the range. Unknown durations remain eligible; manual reprocessing overrides the filter. Configuration > Episode duration filters
Differential hold - An uncorroborated cross-fetch differential candidate: the two fetches measurably differ, but no other stage, overlap, or matched audio cue backs it as an ad. How It Works > Cross-Fetch Differential
Download history - The History page’s Downloads tab: one row per client download of episode audio, with time, client IP, optional location, subscriber label and client app. Repeat requests from one client within 15 minutes merge into one row, and rows are pruned after 90 days by default. Download History
Download User-Agent override - A per-feed download identifier in Feed settings > Advanced. Blank inherits the global string; cross-fetch still uses a different podcast client. Configuration > Outbound Requests
Description chapter trim - A per-feed rule that cuts a span regardless of detection. The span starts at a description timestamp line or a chapter (published or generated) whose title matches a pattern and runs to the end of the episode or to the next chapter. A matching generated chapter queues one recut. Configuration > Chapter trim
Dry-run import plan - The preview a bulk archive-import scan returns before anything is written: every file matched to an episode, every rejected file with a reason, and the publish date each episode would get. Committing re-checks the same files and refuses if anything changed underneath the plan. Local Feeds > Scan, then commit
Enclosure length - The audio file size in bytes declared in an RSS item. MinusPod uses the processed file size after cutting and the publisher’s declared size for unprocessed audio. Podcasting 2.0 > Regenerated for the processed audio
Episode statuses - Discovered (seen in the feed, not processed), Pending (queued), Processing (running now), Completed (processed and in your feed), Failed (will retry), Permanently failed (out of retries), Deferred (waiting on an offline endpoint). How It Works > Processing Queue
Extra yt-dlp arguments - A YouTube setting whose text is split like a shell command and added before MinusPod’s own options on every yt-dlp call. Options that run programs, move files, read local files, change the JSON output or update the binary are rejected. YouTube Feeds > yt-dlp
Failover - A standby LLM provider or transcriber that takes over a target (Provider A, Provider B, or the active transcriber) when it is unreachable, times out, or rejects the key, model, or billing. Switches back after enough consecutive healthy health-probe checks for an automatic switch, or on manual cancel for one triggered by hand. Failover
False-positive text - Transcript text stored from a confirmed “Mark as Not Ad” correction, matched against future episodes of the same podcast to suppress similar text automatically. How It Works > Pattern Learning
Fingerprint - An acoustic signature of a known ad, matched against new episodes without any transcript. One of the pattern types MinusPod learns from confirmed cuts. How It Works > Pattern Learning
Force recheck all - Starts another review of every active learned pattern, including previously reviewed ones, and replaces their pending cleanup suggestions. Work proceeds in batches. Pattern Cleanup > Force recheck
First pass (Pass 1) - The main detection run over the freshly transcribed episode: patterns, cross-fetch, and the LLM reading the transcript in windows. How It Works
Geo lookup - An off-by-default option that sends public client IPs from download history to the ipwho.is service over HTTPS to show a city, region and country. Private and shared-address-space IPs are never sent. Download History > Privacy
Health probe - A periodic check that an LLM or transcription endpoint is usable. Failover uses probe results to detect outages and recover the original provider; fresh results are shared across queued episodes. Failover > Health probes
Held for Review - An ad that detection wanted to cut but a guard stopped. The audio stays intact until you approve or dismiss it. For the causes see How It Works > Held for review; for the actions see Web Interface > Held for Review
Held remainder - The part of a larger detection that stays held for review after a piece of it was cut or released. It inherits the original’s reviewer proposal, which is dropped or clamped when it no longer overlaps the remainder. Web Interface > Held for Review
Hide private episodes - A YouTube setting that drops playlist entries yt-dlp cannot see, such as private or deleted ones, instead of listing them as episodes that cannot be downloaded. On by default, with a per-feed override. Turn it off when the cookies file belongs to an account that can watch them. YouTube Feeds > Private and deleted episodes
Hold (queue) - Any reason the queue is waiting rather than working: a rate-limit pause that stops new claims outright, or an offline-queue wait that parks specific episodes while the rest keep processing. The status bar names which one is in effect, and GET /api/v1/status reports it in a hold block. Configuration > Rate-Limit Hold
Import directory - The user-managed folder for a local feed’s archive import, <data>/import/<slug>/. You place audio and sidecar files there yourself; on a successful commit MinusPod moves the audio out and deletes that episode’s sidecars along with it, leaving a rejected or errored file’s sidecars in place to fix and re-scan. Local Feeds > Bulk import
JIT blocked user agent - An entry in Settings > Security > Agents that skip processing. A client whose User-Agent matches is redirected to the original audio when it asks for an episode MinusPod has not processed yet, instead of starting a run. Download history’s Block action adds ^ plus the client’s first product token. Configuration > Blocked user agents
Keep action - A per-category segment action that leaves a detected span in the audio untouched. How It Works > Segment Categories
LLM fallback - On a YouTube feed, running the normal LLM pipeline when SponsorBlock has no marks after the last probe. On by default, with a per-feed override. YouTube Feeds > How processing works
LLM window cap - The longest single ad the LLM may report within one detection window; a longer one is discarded. The global Longest single LLM detection setting defaults to 7 minutes (60 to 900 s), and a feed’s max ad duration cap or reject override raises it for that feed. Also called the longest single LLM detection. Configuration > Longest single LLM detection
Local feed - A podcast feed MinusPod builds and serves from your own audio files, with no upstream RSS feed behind it. Episodes still run through the same ad-removal pipeline; MinusPod is the publisher rather than a proxy. Local Feeds
Long-form playlist - The channel uploads list without Shorts that a YouTube channel feed reads by default. Turn on Shorts for the feed to read the full list. YouTube Feeds > Shorts and duration filters
Low ad yield - The amber episode badge shown when a run removed far less ad time than the feed’s recent average. Usually a lightly-filled DAI download, occasionally a missed ad worth a look. Web Interface > Processing stats
MP3 stream copy - An opt-in cut path that preserves source-coded MP3 audio outside splice transitions and falls back to re-encoding when the cut is incompatible or fails validation. Configuration > Output audio
Mark action - A segment action that leaves the audio intact and publishes a skippable chapter when chapters are enabled. Keep leaves the audio intact without an ad chapter. Podcasting 2.0 > Ad chapters
Multi-span Mark ad - Selecting several transcript spans under one sponsor and saving each as a separate missed-ad correction. Each span must pass the minimum text length before saving starts. Web Interface > Adding a New Ad
Normalization - A rule that maps sponsor name variants (“betterhelp.com slash pod”, “Better Help”) onto one sponsor so patterns and history stay tidy. Web Interface > Sponsors and Normalizations
Offline queue - An opt-in hold that parks an episode when the LLM provider or Whisper endpoint is unreachable, probes the service every few minutes, and re-queues the episode once it answers. It does not stop the queue: everything not waiting on that service keeps processing. Configuration > Offline Queue
OPML - A portable list of podcast subscriptions. Export all or selected feeds using their original URLs or MinusPod’s ad-free URLs, then import the file into another podcast app. Web Interface > OPML export
Opening window exclusion - A global or per-feed number of seconds at the start of an episode during which ad markers are ignored, so a show’s opening cue is not cut as an ad. Configuration > Detection Tuning
Outbound Requests - The settings section holding the two User-Agent strings MinusPod sends: one for audio, artwork, and chapters, one for RSS. Editable so a host that starts refusing ours can be worked around without a new release. Configuration > Outbound Requests
Probe window - The SponsorBlock lookups made right after a YouTube video is discovered: a set number of attempts at a set interval, after the processing delay. YouTube Feeds > How processing works
Prompt caching - Reusing a static prompt prefix across LLM calls. MinusPod requests a 5-minute cache for Anthropic system prompts and records cache reads and writes inside input-token usage for every provider. Cached input on OpenAI-compatible and OpenRouter calls bills at the model’s cache read rate when one is known. LLM Providers > Prompt caching
Parakeet - NVIDIA Parakeet TDT 0.6B v3, the recommended local transcription backend. It runs in-process through ONNX Runtime and is selected with WHISPER_BACKEND=parakeet. Transcription > Parakeet
Partial detection - An episode published from pattern and cross-fetch cuts alone after the AI detection pass failed. Shows an amber badge and a Re-run detection banner on the episode page; one automatic low-priority re-detect is also queued. How It Works > Partial Detection
Pass 1 detection (YouTube feeds) - A per-feed switch that runs the LLM detector on a YouTube episode in addition to SponsorBlock, even when marks exist. Both sets of segments are cut. Without it, the LLM runs only on episodes with no marks, while learned patterns always run. Full reprocess and Re-detect ads force it for that run. YouTube Feeds > What decides the cuts
Pass-through - One of the five presets on the per-feed Processing mode select. It turns processing off entirely: episodes are downloaded and relayed with no transcription or ad removal, though the audio may be transcoded for serving. The feed URL stays the same, so switching to another mode resumes processing later without touching your podcast app. Configuration > Pass-through mode
Pass-through, per-episode - A flag on a single episode that runs pass-through for that episode alone, even when its feed is in a different processing mode. Useful for one episode a host serves through dynamic ad insertion that should skip ad removal without opting the whole feed out of it.
Pattern - Anything MinusPod has learned from confirmed ads and reapplies to new episodes: text patterns from transcripts and audio fingerprints. Patterns catch repeat ads without spending LLM tokens. How It Works > Pattern Learning
Pattern Cleanup - An experiment that reviews learned text patterns with an LLM and suggests trims, splits, sponsor corrections, category changes, and disabling unused or unreliable patterns. It supports scheduled and manual runs, with a separate provider and model choice. Pattern Cleanup
Podping - An opt-in listener that watches the Hive blockchain for publish notifications and refreshes a matching feed immediately. Its RPC server list is editable in settings and through the API. System Health identifies the active RPC node, checks every fallback every five minutes, and can check them on demand. Scheduled feed polling continues either way. Podcasting 2.0 > Podping
Preparing stage - The first stage a job shows in the status bar, before any download or transcription starts. A run that reuses a retained original, such as a local-feed reprocess or re-detect, stays in Preparing and never shows Downloading. How It Works
Processing mode - The per-feed preset, one select in Feed Settings, that decides what the pipeline does with each episode: standard, keep content only, skip ad detection, pass-through, or cue-only. Episode Processing Workflows > Processing modes
Processing queue - The Queue page shows active jobs and the full waiting list, with controls for queue behavior and priority. How It Works > Processing Queue
Processing stats - The per-run table at the bottom of the episode page: what each run downloaded, detected, cut, held, and verified, plus elapsed stage and FFmpeg timing when available. Web Interface > Processing stats
Provider A / Provider B - The two LLM provider slots a pipeline stage can be routed to. The API also accepts the older primary/secondary slot values and secondary* settings keys. LLM Providers > Per-Stage Providers
Queued - An episode with an actual row in the processing queue, waiting its turn. Not every pending episode is queued: pending only means “not yet processed,” while queued means work is scheduled to start. The episode API’s jobState field reports this directly, rather than leaving callers to infer it from status. How It Works > Processing Queue
Queue priority - A per-feed High/Normal/Low processing-order preference, with automatic boosts for episodes published in the last 48 hours and for manual reprocesses. Configuration > Queue priority
Replacement sound - The clip inserted for Remove actions, enabled by default. Turning it off omits the clip on both copy and re-encode paths; Beep actions still use the replacement clip. Configuration > Output audio
Reasoning budget / effort - Controls for how much thinking an LLM does before answering. Older Anthropic models use a token budget; adaptive-only models use an effort level. LLM Providers > Reasoning settings
Rate-limit hold - An opt-in hold that puts an episode back in the queue when the LLM provider answers a 429 carrying a reset time, and stops the queue claiming any work until that time passes. Unlike the offline queue it pauses everything, Play and Reprocess included. Configuration > Rate-Limit Hold
Recut - Re-cutting the retained original audio using the current ad markers, with no download, transcription, or LLM involved. What “Approve & Recut” does. How It Works > Reprocessing Modes
Reprocess modes - Reprocess (the default, learned patterns plus LLM), Full reprocess (LLM only, called Full Analysis in Configuration), Re-detect ads (reuse the saved transcript), and Recut audio. Configuration > Reprocessing
Remove cue with ad - A cue-template option, on by default, that includes a matched ad-break cue in the cut. Programme cues remain protected. Audio Cues > Remove cue with ad
Resurrected - A detection the validator rejected that the Ad Reviewer overruled and put back in the cut list. Configuration > Ad Reviewer
Retention - How long processed audio is kept before the episode resets to Discovered. The pre-cut original can have its own shorter window, and any feed can override the whole window on its own settings page. Configuration > Per-feed retention
Retired pattern - A pattern disabled after an approved cleanup suggestion because it has gone unused. Its record is retained and the approval can be undone if it has not been changed afterward. Pattern Cleanup > Suggestion kinds
Reviewer proposal (trimmed confirm) - The narrower span the Ad Reviewer suggests for a held detection, kept as reviewer_proposed_start and reviewer_proposed_end. Confirming it approves only that span. The button appears only when the proposal overlaps the held span. Configuration > Ad Reviewer
Search index journal - The search_index_changes table that records every change to searchable content. The search index applies it every 6 hours (refresh) and is rebuilt in full weekly, so a new title or transcript can take up to 6 hours to appear in search. Web Interface
Second scan - See Verification pass.
Seed sponsors - Four toggles, one per LLM pass (detection, verification, reviewer, resurrect), that control whether that pass is handed the known-sponsor list. All on by default; turning one off lets that pass judge each candidate without a prior nudge from sponsors seen before. Configuration > Seed sponsors
Sidecar file - An optional file next to an archive-import audio file, sharing its exact basename: a .txt description, a .jpg/.jpeg/.png cover, or a .json file overriding title, description, publish date, season, and episode. A JSON sidecar overrides everything else, including the filename’s sNNeNN token. Local Feeds > JSON sidecar
Segment category - What kind of content a detected marker spans: sponsor, cross-promo, self-promo, interaction, intro, outro, or recap. Each category resolves to an action (remove, beep, keep, or mark). How It Works > Segment Categories
Silence snap - Nudging a cut boundary to the nearest silence so the edit lands between words instead of inside one. Audio Cues > Silence snap
Sliding windows - Long transcripts are fed to the LLM in overlapping chunks so nothing is missed at chunk edges; the Windows column in Processing stats counts these. How It Works > Sliding Window Processing
Splice check - The rule that holds a long cut for review unless the audio shows an edit point near its edges. Feeds whose ads are never joined into the audio can turn it off on their own settings page. Configuration > Splice check
Splice evidence - A mark in the audio where something was joined: a transition pair, an ad-break cue, a sharp volume step, or a detected splice event. Server-inserted and edited-in ads leave these; an ad spoken straight through in a single take does not. How It Works > Held for Review
sNNeNN naming token - The s01e01-style prefix a local feed’s archive-import files must start with to be matched: case-insensitive, zero-padded to at least 2 digits for both season and episode. Mints the episode’s id and is what sidecar files are matched against. Local Feeds > Naming scheme
Sponsor - The advertiser behind a detection. Sponsors accumulate history per feed, which gets fed back into detection as a hint. Web Interface > Sponsors and Normalizations
SponsorBlock - A community database of timestamps for sponsor reads, self-promotion, intros and similar spans in YouTube videos. On a YouTube feed its marks are the cut source by default. YouTube Feeds > SponsorBlock categories
SponsorBlock state - Where a YouTube episode is in the SponsorBlock gate: waiting, probing, watching, done, exhausted or llm. The UI shows it as one line on the episode. YouTube Feeds > Episode states
Staging area - The per-feed holding folder for a local feed’s uploaded archive-import batch, <data>/import-staging/<slug>/. Unlike the import directory, MinusPod fully manages it: it is populated by the upload endpoint and cleared out as the import commits. Local Feeds > Bulk import
Subscriber feed / subscriber key - A credential scoped to one podcast, labelled for a person or device and revocable on its own, in the form of a 16-character key ID, a dot and a 64-character secret. Download history shows the label of the key that fetched each row. Revoking one leaves the global key and other subscribers working. Security > Authenticated feeds
System One - A structured question protocol used by the native TypeSafe and System One-compatible providers. Detection, verification, and ad categorization are supported; review is experimental. Profiles are independent for each credential slot and provider type. Stats separate logical adapter calls from actual HTTP attempts. System One
Synthesized publish date - The publish date MinusPod assigns a local-feed episode when none was given explicitly: episodes are sorted by season and episode, the newest anchors at import time, and earlier ones step back a day each (or space evenly between two explicit dates). Local Feeds > Publish dates
System One Tuneables - The Settings section for independent native-provider tuning profiles. System One > Independent tuning profiles
Text pattern - A learned chunk of ad transcript matched against new episodes by similarity. Deterministic: if the same ad copy appears, it hits. How It Works > Pattern Learning
Text recurrence hint - An optional signal that flags transcript spans repeating near-verbatim across a show’s last two or more episodes (intros, credits, boilerplate) and passes them to pass 1 detection as a hint. Off by default, and never enough by itself to cut anything. Configuration > Text recurrence hints
Title blacklist - A per-feed list of case-insensitive glob patterns (e.g. Bonus Episode *) that skip processing for matching episode titles. A per-feed choice serves a skipped episode unmodified or hides it from the feed. Manual reprocess overrides it. Configuration > Title blacklist
Transcription fallback - A YouTube setting that lets the configured transcription backend run for videos with no captions. Off by default: such videos are cut from SponsorBlock marks only, or wait for marks. A global setting plus a per-feed override. YouTube Feeds > Transcripts and chapters
Transcript (VTT) - The Podcasting 2.0 transcript MinusPod generates for the processed audio, with cut regions accounted for. Podcasting 2.0
Transcript differential - Comparing the locally transcribed audio with a publisher transcript to find speech absent from the publisher’s copy. Gaps need corroboration before they can be cut. Upstream Transcript Differential
Transcription backend - The engine that turns audio into timed words: faster-whisper, Parakeet, or a remote OpenAI-compatible API. Chosen by WHISPER_BACKEND or in Settings > Transcription. Transcription
User-Agent - The string MinusPod sends to identify itself on an outbound request; there are two, because hosts disagree about what they will answer. Configuration > Outbound Requests
Validation - The rule-based gate every detection passes before cutting: duration limits, confidence, overlap with your corrections, cue evidence, splice checks. How It Works > Post-Detection Validation
Verification conflict - A second-pass finding that lands on a span the first pass deliberately kept. Your segment-action map wins, so the audio is left alone, but the finding holds for review under hold reason verification_kept_conflict rather than being discarded. How It Works > Verification Pass
Verification miss - A standalone pass-2 detection (an ad the verification pass found but the first pass missed) that overlaps no pass-1 marker. Above a confidence floor it holds for review, or cuts automatically when autocut is enabled; below the floor it is dropped and logged rather than surfacing silently. Configuration > Detection Tuning
Verification pass (Pass 2) - After cutting, MinusPod re-transcribes the output audio and runs detection again to catch anything the first pass missed. “Clean” means the second scan found nothing left. How It Works > Verification Pass
Watch window - The days after a YouTube video is published during which MinusPod keeps checking SponsorBlock, queues the episode when marks appear, and recuts it when the marks change. The watch ceiling ends it. YouTube Feeds > How processing works
Waveform Ad Editor - The visual editor for a single detection: waveform, transcript context, and draggable boundaries, with the original audio for reference. Web Interface > Waveform Ad Editor
Whisper (faster-whisper) - The default local transcription backend (WHISPER_BACKEND=local). WHISPER_MODEL and WHISPER_COMPUTE_TYPE apply to it only. Transcription > Whisper (faster-whisper)
Webhook events - Notifications MinusPod can send about episodes, feeds, the queue, services, and failover. Each can also go out by email. API & Webhooks > Events
Webhook template fallback - If a custom payload template cannot render for an event, MinusPod uses a default body. Webhook test results set a templateFallback flag whether or not delivery succeeds. Text content types get a readable line; other types get JSON. API & Webhooks > Webhooks
YouTube feed - A podcast feed MinusPod builds from a YouTube playlist, channel or single video: it lists the videos, downloads the audio with yt-dlp, removes sponsor segments and serves the episodes as RSS. Off by default. YouTube Feeds
Proxy URL (YouTube) - A YouTube setting that routes yt-dlp and YouTube Data API requests through an HTTP or SOCKS proxy. YouTube Feeds > yt-dlp
yt-dlp - The command-line downloader MinusPod uses for YouTube. A pinned copy ships in the image, and a copy on the data volume updates daily. YouTube Feeds > yt-dlp