[antoine-debroye]

Why

Hosts such as Simplecast and Acast insert adverts into the MP3 per request, picked by the listener's country. A French show fetched from the UK carries English adverts; from Germany, German ones. The goal is no inserted advert at all, so detection cannot rely on language, and a proxy only changes which adverts arrive.

The evidence used instead: a host splices in a file it already has, so the same advert is byte-for-byte the same MP3 frames in every episode, while the programme (theme included) is encoded afresh each time.

What changes

• Copied audio (src/lib/copied-audio.js): stretches identical across two or more episodes are cut. New catalogue kind inserted, migration 014.
• Breaks between markers: Acast stitches a 2.2 s clip at both edges of every break. What lies between is cut, even adverts never heard elsewhere. One row per slot (pre-roll, mid-roll n, post-roll), so Restore here restores one break.
• Frame hashes cached beside each fingerprint. Existing episodes are read once, without decoding.
• Second download for feeds whose host inserts (set automatically, or with a new switch on the subscription form): same address, same user agent, a day later. Never a disguise.
• Guards: over 2.5 minutes, or a copy only two episodes share mid-file (a best-of could share programme), is offered rather than cut. Breaks never add up to more than half an episode. An approved cut is never taken away for a row still waiting.
• Jingle fix (bug also in 1.10.1): on Le Rendez-vous Tech the jingle search auto-confirmed the show's own theme and cut the host's spoken intro, up to 1:33. A jingle starting more than 3 s after the host's inserted audio ends is no longer auto-confirmed, and an earlier auto-confirmation is taken back. Jingles the owner confirmed are untouched.
• scripts/check-copied-audio.mjs <folder> runs the search over downloaded episodes. Spec §19.3 and §19.7 and the README are updated.

Measured on real feeds (2026-10-28, from the UK)

Show	Host	Result
Les chroniques d'Arnaud Demanche	Simplecast	13.9 s opener + 10.1 s closer, identical in 30/30 episodes, best-ofs included
Le Rendez-vous Tech	Acast	pre-, mid- and post-roll breaks cut in 5/5 marked episodes; served episodes open on the host's intro
Laurent Gerra	Audiomeans	nothing inserted, nothing cut

Neither Simplecast nor Acast leaves stitch marks, so the old second-download check never ran for these shows.

Testing

• npm test: 1476/1476 (was 1428). Each new rule has a test, and each test was checked to fail with its rule removed.
• test/acceptance/run.sh against a locally built image: 65/65, including a new step 24.
• The real episodes above, run through the built container in automatic mode; transcribed either side of every cut and at the start of the served files.
• Browser check of the Adverts page and the subscription switch.

Test fixtures that model a repeated theme or read now use a variant header option (test/helpers/mp3.js): same sound, different bytes, as a real re-encode is. Otherwise every synthetic repeat would count as inserted.

Known gaps

• Not yet measured: whether Simplecast changes its adverts from one day to the next.
• An advert inserted into one episode only, outside any marked break, and served identically to both downloads is not found by this.
• One untested safeguard: re-looking for a pre-roll when a jingle is taken back late in a pass.

🤖 Generated with Claude Code