kajisho5/ffmpeg-skill v3.7.0

on GitHub

Automated release: version and notes generated from pull requests merged since 3.6.7.

Added

• audio.py --tempo FACTOR and --length-samples N (#486). --tempo changes the tempo of
  the main track and keeps its pitch, with atempo stages of 0.5–2.0 chained for a factor outside
  that range (0.1–10); --length-samples pads with silence or trims the track to N samples per
  channel before it is encoded, and a length_samples verification step decodes the written track
  and counts them: the count must be N for a lossless output (PCM, FLAC, ALAC) and within one
  codec frame of N for a lossy one, whose encoder pads to whole frames. With
  --length-samples, an Opus output counted at a rate other than 48 kHz is refused (an Opus stream
  decodes at 48 kHz, so the file would not hold N samples), with --sample-rate 48000 as the way
  out; without --length-samples an Opus output of any rate is written as before. The result's timing reports {tempo, atempo, length_samples, sample_rate}. Both
  need an audio output (--tempo also takes a --replace track beside the picture): the input's
  own sound beside its stream-copied picture would drift from it, and is refused.
• audio.py --sample-rate HZ and --channel N (#487). --sample-rate resamples the output
  (aresample, after the mix); --channel N keeps one channel of the track by its 0-based index
  as a 1-channel output (two identical sides with --stereo). The output's rate and channel count
  are read back by probe, as the sample_rate and channels verification steps; the audio
  block reports what was asked (sample_rate, channel), not a measurement.
  An index past the track's channels, and --channel with --downmix, are refused before ffmpeg
  runs.

Fixed

• A source tagged colour space gbr can be graded with --keep-alpha (#454). FFmpeg's PNG
  encoder tags a PNG-in-.mov/.mkv stream gbr, and the alpha line passed it back as
  -colorspace gbr, a name the option refuses, so every color.py --lut/--correct --keep-alpha of
  such a source failed (graphics.py and overlay.py share the line). gbr is no longer carried
  over (the PNG line tags itself; a YUV line would be mislabelled by it). The sweep of every probed
  tag passed back as an option found a second name: -color_trc refuses ffprobe's bt470m and
  bt470bg (PAL/NTSC-tagged sources), which failed the same alpha line and made cut.py --smart
  fall back to a full re-encode. _common/color.carried_tag() now spells every such tag the way
  its destination takes it.
• A video-only VP9/AV1 .webm is not refused for an audio encoder it never uses (review of
  #492). The new WebM check (#488, below) read -c:a aac in cut.py's line for a silent source
  as a sound WebM cannot hold; an encoder option is only a reason to refuse when an input has a
  stream of that kind.
• audio.py --limit no longer delays the sound (#476). alimiter ran without latency
  compensation, so the output was the input delayed by the look-ahead (--limit-attack, 5 ms by
  default: 240 samples at 48 kHz) and drifted off a stream-copied picture. It now runs with
  latency=1; so does --voice strong's limiter, the one other alimiter in the skill. The
  limited sound starts on the input's own sample (measured, within 3 samples) and keeps its length.
• audio.py pads and trims to the sample (#478). The sound was conformed to the source with
  apad,atrim=0:{seconds:.3f}, rounded to the millisecond, so a --replace or --music result,
  or a track beside a stream-copied picture, could be up to 24 samples (48 kHz) longer or shorter
  than the source. It is now apad=whole_len=N,atrim=end_sample=N, N counted at the track's own
  rate (pinned with aformat), as join.py has done since #455.
• A .webm output of an H.264/AAC encode is refused before anything runs (#488).
  crop.py -o out.webm failed in the muxer after the encode had started ("Nothing was written into
  output file", kind: ffmpeg) while its dry run said completed. The cause is shared: every
  re-encoding tool writes H.264 (or HEVC) and AAC, which WebM cannot hold, and the sweep of every
  tool in tests/sweep.py with a .webm output failed the same way (24 of 25; look writes a
  PNG). The one check every ffmpeg command passes now refuses it, real run and dry run alike, as
  kind: input with reason output_container_codecs and the encoders in details; name a
  .mp4/.mov/.mkv output instead. VP9/Opus lines (--alpha-codec vp9, export.py --preset vp9-alpha) and stream copies are unchanged.
• feat(audio): --tempo, --length-samples, --sample-rate, --channel; fixes for colour tags, limiter delay, sample-exact length and .webm outputs (#492)

Closes: #454, #476, #478, #486, #487, #488

What's Changed

• feat(audio): --tempo, --length-samples, --sample-rate, --channel; fixes for colour tags, limiter delay, sample-exact length and .webm outputs by @kajisho5 in #492

Full Changelog: v3.6.7...v3.7.0