kajisho5/ffmpeg-skill v3.1.7

on GitHub

Automated release: version and notes generated from pull requests merged since 3.1.6.

Fixed

• silence.py --filler never cut a filler word narrower than about 0.36 s but reported it as
  removed: --margin was taken off each side of the filler span too, which turned it inside out
  (kept_duration longer than the input). The margin is now kept around silences only; a filler
  cut is the word plus --filler-pad. A silence narrower than two margins removes nothing instead
  of overlapping its neighbours (#414).
• Transcripts were early by the sound's start offset in caption.py --transcribe and
  silence.py --filler --transcribe: the engine's 16 kHz WAV began at the sound's first sample, so
  with a sound 0.276 s after the picture every cue landed 0.276 s early and --filler cut the wrong
  276 ms. The WAV (and --karaoke's word-energy decode) is now padded to the file's clock
  (aresample=async=1:first_pts=0, as #409 did for decode_pcm_mono) (#414).
• silence.py --filler --transcribe on silent audio blamed the Whisper build ("produced no
  word-level timings") instead of refusing with reason: "no_speech" as contract.md promises;
  a Whisper engine whose transcript holds no text is now no_speech (#414).
• caption.py --transcribe on a video with no audio stream failed unclassified (kind: ffmpeg,
  "Stream map '0:a:0' matches no streams"); it is refused as kind: input before any engine
  runs, dry run included, as silence.py refuses it (#414).
• render.py --export-timeline (#414):
  • ignored the source's start timecode: a camera file tagged 01:00:00;00 was exported with
    source in/out 00:00:01:00, outside the media for a conform. EDL source timecodes now count
    from it (drop-frame labels where the source's timecode is drop-frame), the FCPXML asset
    start and OTIO available_range start at it;
  • counted EDL source timecodes at the sequence rate, so a 25 fps clip in a 29.97 sequence got
    frame :29; each clip's source timecode is now at its own rate;
  • kept a clip out past the end of its media (10 s of a 6 s clip, where the render clamps);
    the clip now ends at the media's end, with a note, and an in at or past the end is refused
    as cut.py refuses it;
  • dropped a chapter at or past the end silently, while the render's metadata.py refuses it;
    such a chapter, or one that does not come after the one before it, is now refused
    (kind: input).
• caption.py --mode mux copied an input's own subtitle tracks whatever the output container, so
  subrip from .mkv into .mp4 ("Could not find tag for codec subrip") and mov_text into .mkv
  failed unclassified. Existing tracks are now converted per container as export.py converts
  them (#380's rule, now shared in _common/runner.py, with WebVTT for .webm), a track no
  container conversion fits (a bitmap one in .mp4/.webm) is refused as kind: input, and so
  is a .webm mux of a picture or sound WebM cannot hold (H.264/AAC) (#414).
• caption.py handed a caller's SRT that needed no layout change straight to ffmpeg's SRT reader,
  which reads another grammar than the tool: 00:01,000 (no hours) was refused as
  input_unreadable, and 00:00:01,5 was burned at 1.005 s where the tool counted 1.5 s. A file
  whose timing lines are not in the grammar both share is burned and muxed as the copy the tool
  writes (<out>_adjusted.srt; an extra --mode mux track as <out>_trackN.srt) (#414).
• On Windows an explicit --font was resolved by file-name prefix, so installed families were
  missed and reported as font_substituted "not installed here": Malgun Gothic resolved to
  YuGothM.ttc (no Hangul), Microsoft YaHei, Nirmala UI, Leelawadee UI and Times New Roman to arial.ttf. The family's own file is now looked up first in the table of what Windows
  installs under which name (WINDOWS_FONTS, plus the common Latin families) (#414).
• contract --json left out output keys the tools emit and contract.md documents
  (dropped_non_av_streams on 19 tools; cut's mode, requested_*, output_duration,
  duration_delta_seconds, nearest_keyframes, snap; silence's cut_points, cut_warnings,
  removed_seconds_total, timemap, silence_kinds, silence_breakdown, hint; probe's
  timecode, chapters, subtitle_stream_details) and three error reasons the ASR refusals
  emit (no_speech, engine_failed, english_only). They are published, those three refusals
  now also carry error.reason/error.details (the top-level reason stays), and the test that
  lists every refusal reason reads die(..., reason=...) calls too (#414).
• Review follow-ups: caption.py --mode mux reports each track's language/title as the written file carries them, checks the --audio-stream it copies for .webm, converts an EIA-608 track (kept as is in .mov), and no longer copies an SRT whose cue text holds -->; silence.py refuses a non-finite or negative --margin; a Whisper segment with no text is not no_speech; a start timecode the source rate cannot have is read as none, and the EDL M2 speed is at the clip's own rate (#414).
• Subtitles, transcription, timeline export, Windows fonts and contract drift (#414) (#427)

What's Changed

• Subtitles, transcription, timeline export, Windows fonts and contract drift (#414) by @kajisho5 in #427

Full Changelog: v3.1.6...v3.1.7