# Level Desk

> A volume setting is a multiplication factor, not a target level. A file peaking at -24.6 dB
> played at 1.0 is still at -24.6 dB, and with the bed at -9.9 it is fifteen decibels underneath.
> Capping at 1.0 to be safe is not caution, it is silence. Five lanes over one bed-and-cues sheet,
> plus a free browser-side gain-staging engine.

URL: https://level-desk.skillsafe.ai/
API: https://level-desk.skillsafe.ai/api.html

Derived from Vincentwei1021/video-shotcraft (https://github.com/Vincentwei1021/video-shotcraft),
Apache-2.0: its sound-design reference, which states the misconception with measured numbers and
verifies 20*log10 against real renders. Not affiliated with or endorsed by Vincentwei1021.

## The arithmetic

```text
dB(gain)    = 20 * log10(multiplier)        4x = +12dB, 16x = +24dB
level       = measured_peak_dB + dB(gain)
margin      = level - bed_level
needed gain = 10 ^ ((target_margin + bed_level - measured_peak) / 20)
```

Always give the multiplier, what it is in decibels, and the resulting level. "Turn it up" cannot be
checked; "8.5x, +18.6 dB, landing at -6.0 dB" can.

## Three consequences

- The gain a quiet file needs is often above 1.0, and the PREVIEW CLAMPS THERE. The cue plays
  quieter in the editor than in the export, which is the reverse of every instinct about previews.
  Judging by ear in the timeline leads to raising it further and then the render clips.
- Gain lifts the noise floor by exactly as much as it lifts the signal. A file needing +18 dB
  brings its hiss up 18 dB. This is why replacing the asset beats rescuing it.
- Cues landing together sum. Two independent sources at -6 dB are at -3 dB. Two copies of the SAME
  file are at 0 dB, because identical signals add in amplitude rather than power.

## When a file is too quiet, in order

1. Replace it with one that peaks near zero.
2. Pre-normalise it once into the library: `ffmpeg -i in.mp3 -af "loudnorm=I=-16:TP=-1.5" out.mp3`.
3. Carry the gain in the timeline, above 1.0 if needed, knowing the preview lies and the floor
   comes up.

Never cap at 1.0 as a safety measure. It is the failure this exists to prevent.

## Normalisation

Platforms turn a loud master down to their own figure. A master at -8 LUFS on a -14 platform is
turned down 6 dB: it plays at exactly the level a -14 mix would have, with 6 dB less dynamic range.
YouTube, TikTok, Reels, Shorts and Spotify normalise to -14 LUFS; Apple Podcasts to -16; EBU R128
broadcast to -23.

## Bursts

A run of one file at one level is a machine, not a rhythm. Alternate two samples, step the level
down across the run, and when the hits smear together let the run blur into one sound rather than
voicing each - the ear cannot separate them either.

## The lanes

- `bed` - where to put the music, and what range that leaves above it
- `stage` - every cue converted, placed against the bed, checked for the invisible three
- `burst` - stops a repeated run sounding mechanical
- `normalise` - mix to the platform target rather than past it
- `render` - what to measure in the export, since the preview clamps and the render does not

## What it does not do

It does not decode or process audio. It takes peaks you measured and does the arithmetic on them,
so it is only as good as those numbers.