- The real fix behind the last few rounds. Measuring showed the knobs could give
EITHER a clean low cut-count with no quick cuts (v0.94/v0.96 ~560, 0% shots 1-2s)
OR lots of quick cuts but way too many total (v0.95 722) — never the reference edit's mix of
both (476 cuts, 39% quick 1-2s + long monologue holds). Root cause: OnMic decided
what to keep by DURATION, so it couldn't tell a genuine 1.5s turn from a 1.5s laugh.
- New logic decides by CONTENT: when a camera pops in briefly between two runs of
another camera, OnMic now checks whether the previous speaker KEPT TALKING through
it. If they did, it's talkover (a laugh/breath/"mhm") and is suppressed — the shot
stays on the person who has the floor (fixes Wrenn's "switches on a 1-second laugh").
If the previous speaker went quiet, it's a real hand-off and the quick cut is KEPT.
- Because false switches are now caught by content, the minimum shot can drop back to
1.2s so genuine quick turns survive — aiming at the reference edit's bimodal feel. Unit-tested:
talkover collapses, real hand-off is preserved. New knob interject_yield_frac (0.5).
- Even cutting + gap guard (v0.95) and the calmer suppression (v0.96) stay.
- Measured v0.95's actual output vs the reference on the same 47-min podcast (FCPXML,
real switch counts): the reference edit 476 cuts (median 3.0s); v0.95 came out at 722 cuts
(median ~1.8s) — the 1.0s floor let in a flood of micro-cuts, choppier than both
the reference edit and v0.94 (559). Root cause isn't shot length, it's that OnMic fires more
raw switches than the reference edit on borderline vocal events (laughs, breaths, "mhm").
- v0.96 attacks that directly: minimum shot back to 2.0s, hold-through-interjection
2.5s (was 2.0), dominance margin 10 dB (was 9) so a quieter breath/backchannel
no longer wins a switch, switch delay 0.30s (was 0.25). Net goal: fewer, cleaner
cuts nearer the reference edit's 476 and Wrenn's "it switches on a 1-second laugh" report.
- Max hold 22s (was 30) so the occasional variety "wide" can fire again — v0.95 sat
at 0% wide vs the reference edit's 5% because a 30s cap pushed the wide threshold past ~49s.
- Even cutting (every angle cut at every point) and the gap guard from v0.95 stay.
- Exported the reference hand-edit and OnMic v0.93/v0.94 of the SAME 47-min
podcast to Final Cut Pro XML and counted every real (top-most-visible) camera
switch. Measured, not guessed:
Reference: 476 cuts (10.0/min), median shot 3.0s, floor exactly 1.0s,
39% of shots 1-2s (quick banter), 29% held 7s+ (max 67s), wide 5%.
v0.93 : 922 cuts (19.4/min), median 1.3s -> way over-cut.
v0.94 : 559 cuts (11.8/min), median 2.8s -> close on rate, but the 2.5s
floor killed the fast 1-2s exchanges (only ~1%) and the 12s cap
stopped long monologue holds (max 21s vs the reference edit's 67s).
- v0.95 reshapes Balanced to the reference edit's actual distribution: minimum shot 1.0s
(was 2.5; also matches the "ignore cuts under 1s" default you asked for) so
rapid back-and-forth gets its quick cuts, and max hold 30s (was 12) so a
monologue sits on one camera instead of getting chopped. Switch delay stays
0.25s and interjection hold 2.0s, which suppress backchannel twitch so the
1.0s floor gives real turn-taking, not chatter.
- Not yet matched: wide-shot usage (the reference edit 5% vs OnMic ~1%) is a separate lever;
left for a follow-up after confirming the new cadence on your machine.
- EVEN CUTS (customer feedback): OnMic now razors EVERY camera angle at EVERY edit
point, like the reference edit, so every angle has a clip at every cut and you can drop in
and switch any shot by hand. Before, only the live angle was cut and the wide
stayed one long clip, which made manual overrides painful. Still per-video-track
QE razoring, so mic/audio tracks are never touched.
- This requires each camera angle to be one continuous piece of footage across
the edit. If a track has a GAP (missing footage, or footage that starts late /
ends early), Apply now stops before changing anything and tells you exactly
which angle and where (e.g. "V3 at 00:12:34"), so you can close the gap and
re-run. Contiguous multi-file angles (clip A ends where clip B starts) are fine.- Big pace calm-down after real cut-data analysis showed Balanced was flipping
cameras about every 1.4s (60% of shots under 2s). Now: minimum shot 2.5s (was
1s), switch delay 0.25s (was 0.1s), hold-through-interjection 2.0s (was 1s),
overlap window 0.8s. Result: far fewer, longer, more deliberate cuts.
- Reverted the too-aggressive v0.93 dominance change (back to 9 dB).
- Removed the always-on wide "floor" — the wide track now cuts on/off like every
other camera (matches a conventional hand edit).
- A speaker now wins their own solo camera when they're 6 dB louder than light
bleed/backchannel (was 9 dB), so a host who has a solo angle gets cut to when
they talk instead of the edit sitting on the wide. Also lowered the interjection
hold (1.0s, was 1.4s) so ~1s turns get their camera. Net: more accurate per-
speaker switching and less time on the wide.
- Fixed wide shots / switch boundaries flashing BLACK: the bottom camera track
(lowest V-index — your WIDE) is now an always-on base layer that's never cut or
hidden. Higher cameras toggle on top of it, so any moment they're all hidden
shows the wide instead of a black frame. Put your wide on V1.
- Balanced pace is more responsive by default: Delay Cuts 0.1s (was 0.15s) and
Ignore Cuts < 1s (was < 2s), so quick exchanges get caught.
- AI self-heal: if the silero voice model / ONNX runtime is missing (e.g. a
code-only update that didn't carry the large AI files), the panel now downloads
them automatically from the same CDNs and re-enables AI — no reinstall needed.
- Loud guard: OnMic now warns clearly (pre-cut checklist + a confirm before Create
edit) when AI voice detection isn't loaded, instead of silently making a much less
accurate loudness-only edit.
