You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Without an AI key, a highlights video (--extract_mode long) fills the asked length again: when no provider answered, one failed call returned a single fixed 30 s slot (a 33 s reel for a 180 s target); now the signal picks fill it (176 s for 180 s on the NASA sample).
Clips and highlights skip a weak opening: a mic check ("testing, one two three") or a sentence that leans on the one before it ("And, and it just goes ...", "Uh, ...") is skipped at the start (up to 20 s, never below the minimum length), and the no-AI picker ranks such windows lower.
Without an AI key, a title never ends on a trailing-off fragment ("... places that.", "... data about."); the next complete sentence is used.
--layout auto fits the whole picture (centred over a blurred copy) when no face shows on screen in a clip, as in slides or a screen recording, instead of cropping the content off the sides. Explicit layouts are unchanged.
Review fixes: "So I quit my job ..." / "That is the biggest mistake ..." are no longer treated as weak openings (only connectives like "And", "But", "Because", filler sounds, and "So,"/"Well," with a comma are); the opening skip no longer adds up when a clip is snapped twice and never gives up more than 40% of a clip; when two clips overlap the lower-scoring one gives way (the better earlier clip kept losing its payoff sentence); wide shots of people are not mistaken for slides (a smaller face counts in the source check); any unexpected error in the moment picking now falls back to the signal picks instead of one fixed 30 s clip; Hindi/Japanese/other non-English text is no longer penalised by the English opener rules, and Japanese/Hindi sentence ends (。!?।) count.
--tighten keeps loud pauses (laughter, applause, a reaction after a punchline): a gap between words is cut only when it is clearly quieter than the speech; quiet gaps and gaps under steady background noise are still cut.
Clip benchmark: len/target (share of clips within 25% of the asked length; independent of the sentence rules).
Without an AI key, highlights-video chapter titles no longer come out as "I'm, Back, You're": contractions and weak words are skipped, so a chapter gets its content words ("Bone, Density, Loss"). The clip benchmark no longer grades the chapter list (timestamps) of a description as claims.
A clip no longer ends on a sentence that trails off on a word needing a continuation ("... the places that."): it ends on the sentence before, if that costs under 30% of the clip.
Clip benchmark: cue (a question, number or direct address in the first 3 s), self-cont (the first word is not an unresolved "It/This/They ...") and clean end (the last sentence does not trail off).
README: a short overview around the podcast use case (install, what you get, how it is measured, privacy, comparison); the full reference of every mode, option, API and recipe moved unchanged to docs/FEATURES.md. The README recipes no longer use flags that are not implemented.
Repository clean-up: the old hosted dashboard (web/, server.py and its test, the seed script), a stray notes file and an unrelated demo clip were removed; mcp_server.py (the MCP server) is unchanged.
Privacy wording: --visual_context says the frames go to Google when a Gemini key is set; the browser app footer and README list what leaves the PC when you switch it on; the auto layout description mentions the fit case.
Slides and screen recordings that take the fit layout under --layout auto also zoom into the part of the picture that changes (browser tabs and empty margins stay out; at most about 1.7x, same 16:9 window, so the captions stay below the picture). Conservative: only when a content area of at least a quarter of the frame keeps changing; a cursor, sidebar, webcam inset or a screen where everything changes is left as a plain fit.
Clip benchmark: hook ok (does a clip open on a line that stands alone), reel placeholder titles ("Clip #N") are not counted as ungrounded, and the workflow uploads one sample frame per clip.
Browser app: the timeline strip in "Edit this short" has drag handles. Drag the blue start or end edge; letting go fills the Start/End box (1 s minimum between them) and redraws the strip.
Without an AI key: picked shorts are at least 30 s apart (the top-scoring windows could be neighbours, one scene cut in two).
--tighten works for every short again: the shorts are cut in parallel and shared one filter-script file, so ffmpeg read two plans mixed together and most shorts were silently left untightened (4 of 5 in a podcast run). Each short now has its own.
--tighten never cuts a spoken word: on stage recordings (a keynote with applause) the loudness rule took 6-17 s of speech out of every clip, mid-sentence. Now every transcribed word is kept and only what lies between words is cut (speech lost on that keynote: 0).
--tighten also cuts pauses between spoken words (0.8-4 s, a margin kept), so pauses under steady background noise are removed too; loudness alone left them in. On the NASA sample: silent share 4% -> 0%, a 25 s clip lost 3.7 s of pauses. Not applied to game footage; longer gaps (music, action) are kept.
Clip benchmark: gaps s (pauses between words in the picked span) and cut s (what tightening removed, from the finished files).