Grab a clip from a YouTube video
Pick ten seconds of the person talking — no music, no
second speaker, no crosstalk. Ten to fifteen seconds is plenty;
cloning does not improve past that.
yt-dlp -x --audio-format wav \
--download-sections "*1:23-1:33" \
--force-keyframes-at-cuts \
-o narrator.%(ext)s \
"https://youtu.be/VIDEO_ID"
The * marks a time range rather than a
chapter name, and times are MM:SS or
HH:MM:SS. Without
--force-keyframes-at-cuts the cut lands on the nearest
keyframe instead of where you asked — it re-encodes, so it is slower
but lands on the second.
Or download the whole thing and trim it here instead.