The WeaveSignal contributors
466b32ab6a
The first public release. What it does: https://weavesignal.top; what changed: CHANGELOG.md.
81 lines
3.7 KiB
Text
81 lines
3.7 KiB
Text
# SPDX-License-Identifier: AGPL-3.0-only
|
|
# SPDX-FileCopyrightText: 2026 The WeaveSignal contributors
|
|
#
|
|
# The end of playout's chain (ADR-0044's "one processing chain"), in one place: weavesignal.liq
|
|
# includes it for the programme, and tests/playout/chain.liq and nan.liq include it to test the
|
|
# chain that plays (review 0028, C1: they used to test a hand-made copy).
|
|
#
|
|
# A conservative chain: a slow automatic gain control towards the station's loudness target, then
|
|
# a peak limiter. It evens out what the desk does not process (the library, the fallback files,
|
|
# the default loop) and leaves programme at its target alone: a signal already at the target
|
|
# gets no gain. Single band on purpose, so there is no tone to tune and nothing that pumps on
|
|
# speech.
|
|
# - the target is in LUFS (a 3 s window), -16 by default: the operator's choice of 2026-11-02
|
|
# after a listening comparison against -23 and off (research/2026-11-02-loudness-comparison.md
|
|
# in the project folder). Every source WeaveSignal plays measured -16 to -19 LUFS, so -23 mostly
|
|
# turned the station down; -16 is near where internet streams sit;
|
|
# - the gain stays within +-max_gain dB (6 by default), rises over about 15 s and falls over
|
|
# about 2 s, so a loud track is brought down quickly and a quiet one up slowly;
|
|
# - under -50 dB it holds its gain, so a quiet room or the desk's silent bed is not pumped up.
|
|
# A clip comes before and after the gain control: once NaN reaches `normalize` or `compress`,
|
|
# their state stays NaN and everything after is silence until playout restarts (staging,
|
|
# 2026-11-01); `clip` maps NaN to 0 and leaves real audio alone.
|
|
|
|
# The loudness target when the station's data gives none.
|
|
default_target_lufs = -16.
|
|
|
|
# `gain` is a reference that is given the gain control's own gain getter (for
|
|
# `processing.status`); it is left alone when the gain control is off.
|
|
def weavesignal_chain(~enabled=true, ~target=default_target_lufs, ~max_gain=6., gain, s) =
|
|
s = clip(s)
|
|
s =
|
|
if
|
|
enabled
|
|
then
|
|
agc =
|
|
normalize(
|
|
id="agc",
|
|
lufs=true,
|
|
target=target,
|
|
up=15.,
|
|
down=2.,
|
|
gain_min=-max_gain,
|
|
gain_max=max_gain,
|
|
threshold=-50.,
|
|
window=3.,
|
|
track_sensitive=false,
|
|
s
|
|
)
|
|
gain := agc.gain
|
|
(agc : source)
|
|
else
|
|
(s : source)
|
|
end
|
|
s = clip(s)
|
|
# A peak limiter as the last guard (the desk already limits its programme at -1 dBFS, carts
|
|
# included): peak mode, 1.5 dB under full scale, then a clip that catches what a 1 ms attack lets
|
|
# past. Below its knee it does nothing.
|
|
# `lookahead` must stay 0 in Liquidsoap 2.5, in peak and RMS mode alike. With any lookahead, `compress`
|
|
# keeps its detector's samples in a ring buffer made with OCaml's `Array.create_float` (mm's
|
|
# `Audio.create`), which is uninitialised memory; the first `lookahead` samples it reads are
|
|
# whatever the heap held. A NaN or an infinity there makes its gain NaN for good on the first
|
|
# frame, and the clip after turns that into zeros until playout restarts; the clip before cannot
|
|
# help, the input is clean. Measured 2026-11-02 (real time, this script with lookahead 5 ms and
|
|
# silence in): 7 of 80 starts silent for good, 0 of 80 with lookahead 0; q-all's script, without
|
|
# the gain control, 0 of 240; a bare sine on an unsynced clock, 133 of 300 against 0 of 300. The
|
|
# rate is whatever the heap holds, so any lookahead is unsafe. It bought nothing either: the
|
|
# buffer delays the detector, not the audio, so the first 5 ms of a peak passed untouched.
|
|
s =
|
|
compress(
|
|
id="limiter",
|
|
attack=1.,
|
|
release=150.,
|
|
ratio=20.,
|
|
threshold=-1.5,
|
|
knee=1.,
|
|
lookahead=0.,
|
|
window=0.,
|
|
s
|
|
)
|
|
clip(s)
|
|
end
|