ELEVENLABS INPUT — SYNC TEST CASE 01
=====================================================================
Recommended voice settings for documentary narration:
  Model:       Eleven Multilingual v2
  Stability:   0.45
  Similarity:  0.75
  Style:       0.15
  Speed:       1.0

Render these as FOUR SEPARATE FILES. Do not generate one file with
long silences in it - break tags are unreliable past a few seconds
and you need frame-accurate gaps for a sync test.

---------------------------------------------------------------------
FILE 1  ->  vo/vo_seg02.mp3      (target 8.8s, hard ceiling 9.0s)
---------------------------------------------------------------------
Twelve months later it was gone. The Soviet Union did not collapse in a single afternoon - it came apart republic by republic.

---------------------------------------------------------------------
FILE 2  ->  vo/vo_seg03.mp3      (target 8.0s, hard ceiling 9.0s)
---------------------------------------------------------------------
Through the winter of 1991, the red flag over the Kremlin was lowered for the last time on December twenty-fifth.

---------------------------------------------------------------------
FILE 3  ->  vo/vo_seg04b.mp3     (target 3.2s, hard ceiling 3.8s)
---------------------------------------------------------------------
Not slightly wrong. Obsolete, from the first day.

---------------------------------------------------------------------
FILE 4  ->  vo/vo_seg05.mp3      (target 8.4s, hard ceiling 8.8s)
---------------------------------------------------------------------
Cartographers spent a decade redrawing borders that had held for seventy years. Some of those lines are still argued over today.

=====================================================================
ASSEMBLING THE FULL VO TRACK (narration.wav)

The full track is 45.0s with silence where on-camera dialogue plays:
silent 0.0-9.0 (clip 1) and 27.0-32.2 (segment 4A).

Pad each VO file to its exact slot, then concatenate:

  ffmpeg -y \
    -f lavfi -t 9.0 -i anullsrc=r=48000:cl=mono \
    -i vo/vo_seg02.mp3 \
    -i vo/vo_seg03.mp3 \
    -f lavfi -t 5.2 -i anullsrc=r=48000:cl=mono \
    -i vo/vo_seg04b.mp3 \
    -i vo/vo_seg05.mp3 \
    -filter_complex \
      "[1:a]apad=whole_dur=9.0[a2]; \
       [2:a]apad=whole_dur=9.0[a3]; \
       [4:a]apad=whole_dur=3.8[a4]; \
       [5:a]apad=whole_dur=9.0[a5]; \
       [0:a][a2][a3][3:a][a4][a5]concat=n=6:v=0:a=1[out]" \
    -map "[out]" -ar 48000 -ac 1 narration.wav

Slot budget, which must total 45.0:
  9.0 (silence, clip 1) + 9.0 (seg2) + 9.0 (seg3)
  + 5.2 (silence, seg 4A) + 3.8 (seg4b) + 9.0 (seg5) = 45.0

Verify:
  ffprobe -v error -show_entries format=duration -of csv=p=0 narration.wav

If any VO file renders longer than its ceiling, regenerate it at a
slightly higher speed setting. Do NOT shrink the following gap to
compensate - that is exactly the cumulative drift this test is meant
to catch, and hiding it here defeats the purpose.
