Skip to content

keryx voice takes

Description

See what's on the table before you pick. voice takes lists a workspace's VO candidate takes (vo/takes/NN-T.mp3) per storyboard line, with each take's duration and a marker on the promoted one — so the "re-roll these lines and choose the best" loop can be screened without leaving the CLI (spec 0023). Actual listening stays out-of-band (or in the studio); keryx surfaces the metadata that ranks them.

It is one half of the take workflow:

keryx voice gen --workspace my-post --lines 3,5,7 --takes 3   # re-roll a subset
keryx voice takes --workspace my-post                          # screen what came back
keryx voice pick 5 2 --workspace my-post                       # promote the winner

Usage

keryx voice takes --workspace <slug> [--line N] [flags]

Behaviour

  • Per line: every candidate take with its number, filename, duration, and whether it is the promoted selection (content-matched against vo/NN.mp3 — promotion copies the take, so equality identifies the winner).
  • Durations are best-effort: they probe through the backend selected by providers.render (in-memory on afmpeg); if no backend is available or a probe fails, the listing degrades to 0.00s rather than erroring.
  • --line N filters to one line; --output json emits the listing for the studio / scripts.
  • An empty workspace lists nothing and says so — not an error.

Stilt screening (--screen)

--screen scores each take so a stumble ranks last automatically instead of by ear (spec 0023 R-GEN-38). The combined score is:

  • duration deviation from the line's median take — a longer read (the tell-tale of a stumble or an inserted pause) is penalised more than a tighter one; plus
  • excess internal silence — silences inside the read (lead and tail excluded) beyond the pauses the line actually scripts (… and <break> are forgiven).

Lower is cleaner; the cleanest take is marked ←. The per-take sub-scores are surfaced (in the table and in --output json) so a ranking is auditable, not a magic number. When a line's best take still reads over-paused, the whole line is flagged "all takes over-paused — reword/respell" (R-GEN-38b): no clean read exists, so the fix is the script, not another re-roll.

Silence detection needs the render backend's silencedetect (the default ffmpeg backend); on a backend without it (afmpeg's lean module), screening degrades to duration-only ranking rather than failing.

Flags

Flag Description Default Required
-w, --workspace reel workspace slug yes
--line list only this storyboard line (0 = all) 0
--screen stilt-score the takes (duration + internal silence) and flag over-paused lines false

Run keryx voice takes --help for the authoritative, always-current flag set.

Example

$ keryx voice takes -w my-post
line  take  duration  selected
1     1     3.41s
1     2     3.28s     *
3     1     4.02s

$ keryx voice takes -w my-post --screen
line  take  duration  silence  score  best  selected
1     1     3.41s     0        0.040
1     2     3.28s     0        0.000  ←     *
3     1     4.02s     2        1.290
3     2     3.10s     0        0.000  ←

See also