Repository navigation
Stable release covering 2.102.0 through 2.102.9. The previous stable was 2.101.34.
Breaking changes
- Audio fingerprint matching now follows the pattern's scope. Every stored fingerprint used to match in every feed. Now a podcast-scope pattern's audio matches only in its own podcast, a network-scope pattern's in feeds of its network, and a global pattern's everywhere. An ad that was cut on other feeds through audio alone stops being cut there. The pattern dialog's Audio tab says how many past matches came from episodes outside the pattern's scope. Widen the scope to keep cutting on those feeds (#998).
- Patterns with network scope now match. They never matched before, by text or by audio. If you have network-scope patterns, review them on the Patterns page, because they begin cutting on other feeds of the same network after this upgrade.
PUT /api/v1/patterns/{id}answers 400 when a pattern is moved to network scope with no feed network and nonetworkId(#998). - A path under
/api/that no route matches now answers 404 with a JSON body, and a method the path does not accept answers 405 with JSON and anAllowheader. Both used to return the web app's HTML page. API clients that parsed that HTML need to handle the JSON error (#998).
New features
- Audio fingerprint matching is reworked. The matcher measures the whole ad and reports its measured start, end, similarity and coverage. The old one compared the first 10 seconds and assumed the end from the stored length. A short dropout inside an ad is bridged, and one ad from a stored break of several is found on its own. Audio that shares a music bed with the show no longer matches elsewhere in the episode. How It Works describes the limits on heavily compressed or damaged audio (#998).
- Each pattern's audio is visible and controllable. The Patterns table has an Audio column and an Audio filter. The pattern dialog has an Audio tab with an on/off switch, Play, match counts, a Recent matches list with Not this ad, Edit edges and Delete audio. A pattern with no audio offers Add audio to pattern when a retained episode still holds the ad. Routes:
GET/PUT/DELETE /api/v1/patterns/{id}/audio,POST /api/v1/patterns/{id}/audio/rebuild,GET /api/v1/patterns/{id}/audio/matchesandGET /api/v1/patterns/{id}/audio/preview(#998, #999). - Audio fingerprints are stored from more places, with any LLM provider: Add ad, a confirm or adjust correction, a pattern split and a Pattern Cleanup trim. Audio too plain to ever match, such as silence or a steady tone, is refused with the reason
not_distinctive. Once you delete a pattern's audio, automatic detection does not create it again (#998). - Settings > Audio fingerprinting has five settings: Enable audio fingerprinting, Match strictness, Minimum coverage, Gap tolerance and Minimum ad length. Stats has an Audio matching section, also at
GET /api/v1/stats/audio-matching(#998). - An LLM is optional. Settings > LLM provider has a None choice, also available as
LLM_PROVIDER=none. Episodes still transcribe and run audio cues, audio fingerprints, text patterns and the cross-fetch differential, and nothing calls a model. A regular podcast feed cuts a new ad only after you mark it once, so on a fresh install start with Add ad. See Running without an LLM (#998, #994). - Remove short host asks and sponsor thanks: a new setting, off by default, global or per feed. When it is on, the model may flag a 2 to 40 second subscribe or rating ask, a plug for the host's own work, or a sponsor thank-you. Each flagged marker follows the action set for its category and is cut on sentence edges. API:
snipShortAsks(#1006, #962). - Held for review markers can be widened by hand. "Widen held ads by (seconds)" under Settings, Global defaults (0 to 120, default 0) lets the held editor move a held marker's edges that far past its reviewed span. The editor shows the allowed range and has a "Save, cut later" button (#1008).
- Processing order for queued episodes: As queued, Oldest first or Newest first, set globally under Queue > Queue control or per feed. A back catalog added in one go can now be processed from episode 1 (#998, #997).
- Settings > Security has Session length (hours) and Remember me length (days), and the sign-in page has a Remember me on this device checkbox. Both are also read from
SESSION_LIFETIME_HOURSandSESSION_REMEMBER_DAYSon first start (#998, #996). - Waveform peaks are cached next to the retained original, so the ad, cue and split editors no longer decode the whole file on every open. Episodes longer than 4 hours now have a waveform (#998, #995).
- Episode lists load 256 px cover thumbnails that the browser keeps for 7 days. A cover shared by many episodes is stored once, and cover requests have their own limit of 3000 per minute.
GET /feeds/{slug}/episodes/{episode_id}/artworkacceptssize=thumb(#999, #1000). - Pattern Cleanup suggests removing a sponsor that a pattern's own text never names. The check needs no model call (#1024, #1023).
- Episode webhook payloads carry
podcast.auto_process, which isfalsewhen the podcast is processed on demand (#1025). - A 5-star rating can be stored without confirming cuts or dismissing held markers. Untick "Confirm all cuts and dismiss held ones", or send
confirm_markerson the rating route (#1026, #1007).
Fixes
- Credentials no longer reach logs, stored errors, API responses, webhooks or email alerts. A failed upstream fetch could put a full URL with its query string, such as
key=..., in the error text. Error text stored by earlier versions is cleaned by a background step after startup (#998). - Automatic learning no longer saves an ordinary word as a sponsor. A new one-word sponsor name is accepted only when the ad text ties it to a brand through a domain, a spoken "dot com" or a promo code that contains the name. A sponsor is promoted to a global pattern only from patterns whose own text names it (#1024, #1023).
- A Whisper repetition loop is no longer read as transcript. Text-pattern matching and the validator skip runs of 8 or more short, near-identical cues. A marker that rests mostly on loop text is trimmed, held with the new reason
transcript_loopor rejected (#1005). - A learned text pattern no longer starts a marker one transcript segment early or ends it one segment late. An early start could cut the host's last sentence before a break (#1028).
- An auto-learned text pattern keeps an outro that ends on the sponsor's name alone. A read in a language other than English lost its outro, and the estimated end could reach past the ad (#1030).
- When a text pattern matched only its intro, a break cue right after the read stayed in the audio. The cut edge now snaps onto that cue (#1029).
- An ad found in the second pass is cut when its span only touches a first-pass cut. It was held for review as a reviewer contradiction (#1027).
- Retry ad detection uses the provider chosen for the detection stage. With detection on the secondary provider, every window failed (#1027).
- The pre-roll detector finds the show's opening line in feeds whose title has a subtitle, and the reviewer's inward trim of a DAI pre-roll is accepted when it lands on the edge the transcript differential confirmed (#1004).
- A blank system prompt no longer sends an empty system message. Some gateways rejected it with a 400 (#1003).
- Pattern Cleanup trims and splits update the pattern's audio and stored duration, including for audio stored by earlier versions. A trimmed pattern could still cut its old, longer span (#998, #1024).
- Recut no longer applies a stored adjustment that is under 1 second or that a newer confirm replaced. An adjust that is shorter than 1 second, reversed or outside the episode is refused with a 400 (#1008).
- The Detected ads and Ad review lists could take 10 seconds or more to load on a large database.
GET /api/v1/stats/addressingand the Recents feed query no longer scan every episode (#1027, #998). - The database write-ahead log stays bounded at 64 MiB, and restarting no longer rebuilds the whole search index (#998).
- Memory freed by episode processing is returned to the system once the queue is empty (#999, #1001).
- Parakeet no longer checks the network on each model load. It reads the local copy and checks for a newer revision in the background at most once every 7 days (#998).
- Upstream chapters, transcript and artwork fetches retry once with the feed's user agent when a host answers 403 (#998, #999).
- The Needs review filter and its count on Ad review cover only markers held for review. Detections the validator left uncut are under Not cut.
status=needs_reviewonGET /api/v1/detectionsfollows the same rule (#1027). - On phones, the Ad review and Detected ads cards show their actions as one row of icon buttons, and Confirm is green. The Patterns and Sponsors pages show cards below 1280 px wide, so nothing is clipped or scrolls sideways (#1027, #998).
- A cut row on the episode page says what your confirm or reject did, and a reject offers Recut now (#1026, #1007).
- Every rate-limited 429 carries a
Retry-Afterheader, and the sign-in page says "Too many sign-in attempts. Try again in N minutes." in place of "Invalid password" (#998). - Quieter logs: several repeated warnings now log once or at INFO, and a freshly published episode whose audio is not available yet logs a warning while it has retries left (#998, #999).
Upgrade notes
- Review the three breaking changes above before upgrading if you use network-scope patterns, rely on audio matches across feeds, or call the API from scripts.
- One-time steps run after the first start. Match history is filled in from stored markers, two marker indexes are built, old error text is redacted and the search index is rebuilt once. They run in the background or on first load and delete nothing. Until they finish, Add audio can miss older episodes and the first load of the review lists can be slow.
- A session last used longer ago than the session length (24 hours by default) signs in again once. Changing the password now signs out every device (#998).
- Both compose files now pass
SESSION_LIFETIME_HOURS,SESSION_REMEMBER_DAYS,LOG_FORMATandLOG_LEVEL. Add them to your own compose file if you want to set them. WithLOG_FORMAT=json, gunicorn's lines are JSON too and every line carries alevelfield (#998). - Both images set
MALLOC_ARENA_MAX=2andOPENBLAS_NUM_THREADS=1by default. Override them in compose if needed (#999, #1001). - gunicorn 27 answers a request whose
Transfer-Encodingdoes not end inchunkedwith a 400 (#1024). - A stored reviewer prompt that still says fingerprint spans are projected lengths has that sentence replaced once at startup. Custom prompts are otherwise left alone (#998).
- The "found by audio only" count in Stats can drop after the upgrade, because matches whose stored marker does not say which stages found them no longer carry the badge (#1027).