Upstream Transcript Differential

What it is

Some publishers ship a podcast:transcript tag on their RSS items whose transcript was produced before the ad reads were spliced in, so it is already ad-free. The upstream transcript differential diffs MinusPod’s own Whisper transcript of the downloaded audio against that publisher transcript: any block of speech in the Whisper transcript that the publisher’s copy omits is a candidate ad.

This is an additional evidence signal on top of the detector, not a replacement for it. It never cuts on its own: a gap it finds either corroborates a detection another stage already made, or becomes a new held-for-review marker for a human to confirm. It is best effort, so a fetch or parsing failure never fails the run; the episode processes as if the stage had not run.

Pass-through episodes skip the diff: MinusPod stores the upstream transcript as the episode transcript instead.

Which episodes qualify

• The feed must carry a podcast:transcript tag MinusPod can read: text/vtt, application/srt, application/x-subrip, application/json (podcast namespace transcript JSON), text/plain, or text/html. When an item lists more than one, MinusPod keeps the best type in that order.
• Local feeds have no publisher transcript to compare against and always skip the stage.
• The per-feed setting overrides the global default. Inherit follows the global setting; On enables the stage for that feed even when the default is off.
• The publisher’s transcript URL is read-only evidence. MinusPod never serves it to subscribers; see Podcasting 2.0 > Regenerated for the processed audio.

How a gap is found

Both transcripts are tokenized and aligned word by word. Runs of Whisper-only words, at least 40 words and at least 10 seconds of audio, become a gap span; gaps within 3 seconds of each other merge into one. Below 60 percent word coverage between the two transcripts, the comparison is treated as unreliable (wrong episode, a heavily re-edited transcript, or a different language) and produces no spans at all.

When the publisher’s transcript carries timestamps, MinusPod also checks whether its time axis jumps across the gap by roughly the gap’s own length, the signature of a block that was physically removed rather than just summarized differently. A gap with that signature is marked offset_confirmed and carries a higher confidence (see below); a gap without it, or found in an untimed transcript, still counts as evidence but at the default confidence.

The alignment only ever reads text; it never changes what gets cut by itself.

How a gap is used

Each gap becomes a held marker: category sponsor, confidence 0.6 (0.75 when offset_confirmed), reason “Upstream transcript omits this span”, and hold reason transcript_differential_unreviewed. It never auto-cuts.

• If a fingerprint, text pattern, audio cue, or LLM detection covers at least half of the gap, the two merge and the held gap marker is released. The merged marker spans both and keeps the other detection’s stage.
• The validator also checks every other detection against the gaps. When a gap covers at least half of a detection’s own length, that detection is marked corroborated_by: transcript_differential and its confidence rises by 0.1 (0.15 when the gap was offset_confirmed, capped at 0.95, never lowered).
• If no detection releases it, the gap stays its own held marker on the episode’s Held for Review list, with the hold reason above, until you confirm or reject it.
• Rejecting a held gap marker clears it from that episode without seeding a cross-episode false-positive match: the span is real show content the publisher’s transcript happened to omit, not a confirmed false positive from a real detector, so it should not suppress a genuine future catch on another episode.

The stage runs after transcription and before detection on every run that performs detection, including a re-detection. A re-detection normally re-fetches the publisher transcript fresh. The one exception: a reprocess that reuses the previous run’s saved Whisper segments (rather than re-transcribing) also reuses a stored ok result for the same URL if it is less than 24 hours old, since the word times it would realign against have not moved. Any reprocess that re-transcribes always fetches fresh.

A fetch failure, a parse failure, or an alignment that lands below the coverage floor all leave the run stats and the episode’s upstreamTranscript field with a status (error or unreliable) instead of blocking anything else the run does.

Settings

• Compare with the publisher transcript (Settings > Ad Detection) is the global default, on by default.
• Each feed’s Advanced settings has its own Transcript diff select (Inherit / On / Off), next to Cross-fetch diff, that overrides the global default for that feed. See Configuration > Ad Detection Settings.

The API uses transcriptDifferentialEnabled on PUT /api/v1/settings/ad-detection for the global default and transcriptDifferential on PATCH /api/v1/feeds/{slug} for the feed override. Send true for On, false for Off, or null to inherit on the feed. GET /api/v1/feeds/{slug}/episodes/{id} returns the comparison result in upstreamTranscript, including status, coverage, and detected gaps. The full schemas are in openapi.yaml.

Limits

• Only the Whisper transcript is diffed; there is no attempt to use the publisher’s transcript as a replacement for Whisper, and it is never served to subscribers.
• No speaker diarization: cues and words are compared as plain text.
• The stage runs at processing time or at a re-detection. If a publisher posts their ad-free transcript after MinusPod already processed the episode, nothing re-triggers a recut automatically; a manual re-detection picks it up.