-
Transcribe every audio track of a multi-track file, not the one ffmpeg picks (#1523)
frostbyte_neo 发布于
2026-09-04 13:27:03 +00:00 ffmpeg's default stream selection keeps one audio track, the one with the most channels,
first on a tie. A recording with a track per source (OBS, Discord) lost every other
speaker, and while someone talked on a dropped track the model heard silence and looped
on a line it had heard before: minutes of the same sentence in the transcript. Count the
tracks with ffmpeg's own probe and mix them with amix when there is more than one, at
their own levels and for the longest track's duration; a single track takes the same
command as before, byte for byte. fixtures/two-tracks.mp4 is multi.wav split in two
tracks; the shipped server repeats one line over its silent half, this transcribes it.Closes #1518
Claude-Session: https://claude.ai/code/session_01TgmkaiiuPu65pSqXptMAna
下载附件