Skip to content

Lull 1.11 — push through the quiet parts

Choose a tag to compare

@PortableDiag PortableDiag released this 15 Sep 02:29
· 9 commits to main since this release

The other half of the dead-spot problem. 1.10's Force Tempo sets a speed per track, and that is genuinely all it can do about a quiet bridge: the piano break in the middle of a 140 BPM song is also at 140 BPM, so there is no tempo there to correct. What is wrong with it is that the energy drops out — a different measurement, and a different fix.

Push through quiet parts finds those stretches and speeds up further while you are in one, then eases back out. Under Force tempo in the overflow menu:

  • How much faster in a quiet part — +10% to +150%, default +50%. It goes much further than the whole-track ceiling on purpose: this applies to the passage you specifically want to be over, where stretch artefacts matter far less than the seconds do.
  • How quiet counts as a dead spot — 3 to 24 dB below the track, default 9.

Relative to the track, never absolute. A quietly mastered song is not one long dead spot and a loud one is not free of them, so "full level" is a high percentile of that track's own seconds — not its peak, because one cymbal crash would raise the bar for the whole song and turn ordinary verses into dead spots.

Three guards against speeding up where it would be vandalism:

  • A stretch must last 6 seconds. The bar of held breath before a drop is the opposite of a dead spot, and speeding through it would wreck the one moment the track was building to.
  • Two quiet stretches split by a short fill are one dead spot, not two.
  • Every second of a reported section meets the threshold on its own.

It ramps, and it ramps early. ~800ms, starting 800ms before the quiet part — so it has finished arriving by the time you get there, rather than playing the first and most skippable seconds at the old speed.

One decode, two answers. Where a track goes quiet needs the whole file, unlike its tempo. TrackAnalysis decodes end to end once and produces both — once per track, on a background worker, only while the feature is on, and a track ahead of the one playing. What is stored is the loudness curve, not the conclusion, so moving the threshold slider re-reads every track already measured instead of invalidating a library, and the panel shows live what the current setting finds in what is playing.

API. GET /api/tracks/{id}/quiet lists what was found and distinguishes never analysed from analysed, nothing there; pushQuiet, quietBoost, quietThresholdDb on /api/settings; boosted on /api/player, whose speed is now read off the engine rather than recomputed — it moves during a track now.

Tests: 138 → 171. One is a regression caught on a device: a first cut widened the threshold on the way out of a section, in decibels, which let a run continue through seconds that did not qualify at all — a 4-second dip swallowed the remaining 160 seconds of the track. Gap-merging already gave the anti-flap property that was meant to buy, so the hysteresis is gone.

Verified on device: a generated track with a 30-second quiet section at 1:00 is found at exactly 60000–90000ms at every threshold, the speed ramps 1.181 → 1.772 across the boundary and back, a 4-second dip is correctly ignored at every threshold, and the same holds on the signed release APK.

See CHANGELOG.md for the full entry.