Keep 0.24 s of breath: shorten pauses instead of deleting them

Deleting every gap makes speech sound clipped. Instead pad each kept range by a hold of 0.12 s before and after, so every pause that you cut still leaves 0.24 s. Nine pauses totalling 6.35 s then lose 4.19 s, and each join keeps a short breath.

Instagram says First Draft cuts the pauses and trims your clips, with everything reversible (read 2026-10-09). It publishes no numbers, so any hold value is your own choice.

The arithmetic

A gap of g seconds between two words, with a hold h on each side, leaves min(g, 2h) and removes max(g - 2h, 0). With h = 0.12 a 0.35 s gap loses 0.11 s and a 2.0 s gap loses 1.76 s.

Over a whole clip

Take nine gaps of 0.4, 0.6, 0.6, 0.9, 1.2, 0.5, 0.35, 0.8 and 1.0 s, which sum to 6.35 s. Deleting them all removes 6.35 s; keeping 0.24 s at each of the nine joins removes 4.19 s (6.35 - 9 x 0.24 = 4.19). The same Reel runs 2.16 s longer and sounds more natural.

def pad(ranges, hold=0.12, duration=None):
    """Widen each kept (start, end) by hold on both sides; clamp and merge."""
    out = []
    for s, e in ranges:
        s, e = max(s - hold, 0), e + hold
        if duration is not None:
            e = min(e, duration)
        if out and s <= out[-1][1]:
            out[-1] = (out[-1][0], e)
        else:
            out.append((s, e))
    return [(round(s, 3), round(e, 3)) for s, e in out]

Notes

• The merge step matters: a gap under 0.24 s overlaps after padding and becomes one range, so it is not cut at all.
• STT word times can run slightly tight to the audio. The speech-to-text sentence segments[] option has a boundary_lead_ms setting (default 70 ms, range 0 to 500) for sentence starts; check the inspect docs before relying on it for word padding.
• A longer hold on your padded ranges raises each trim by the same amount, so the 0.2 s trim floor is easier to clear.
• Pick one hold per Reel; changing it re-cuts everything.