Loudness research: measurement, automatic balance and true peaks

Technical research · edition 2 · reviewed 2026-10-24

Prismal Labs native editor capture, illustrating the corresponding software project. View full-size image

What we found

The loudness research separated measurement from control policy. A BS.1770 reading can be correct while an automatic mixer still chases rests, overreacts to transients or overspends its power budget. We therefore treated source qualification, fixed/manual contributions, weighted allocation and gain movement as independent design decisions.

The reusable finding is a normalized power allocator with explicit silence and crest history. It feeds a separate final audio-path meter/limiter. Discovering a Child over the control bus does not route that Child's audio through the Master or prove anything about the final sum.

Allocate the remaining power, not a separate target per track

Every 100 ms, the owning Master reads at most 64 fixed-size snapshots. A source is active when its 400 ms K-weighted Momentary Loudness is at least the configurable Activity Threshold (-55 LUFS by default) and its heartbeat is current. Fixed manual or bypassed rows consume measured post-gain power first. For the remaining automatic channel weights b_i in dB, define w_i = 10^(b_i/10), W = sum(w_i), and let T_rem and C_rem be the target-loudness and peak-power budgets left after fixed rows. For each active automatic channel:

With every weight at 0 dB, this reduces to -10 log10(N). A positive weight moves power toward a featured source and away from its peers; normalization by W means the predicted aggregate target does not rise. A single active source is unchanged by its absolute weight. Active manual overrides are measured after their fixed gain and consume that power first; automatic channels share only the remaining target and peak budgets. The stored weight of a manual channel takes effect again only when that channel returns to automatic mode.

The desired result is clamped to -24 dB and the configurable Maximum Boost (+12 dB by default). Boost waits 1.2 seconds of activity and moves at the selected Response rate; normal cuts move at max(12 dB/s, 4 x Response), and a predicted ceiling violation cuts immediately. Each Child publishes a fast activity flag from its unsmoothed block RMS, so a gap enters hold without waiting for the 400 ms loudness window to decay. Once open, the loudness detector closes 3 LU below its opening threshold and holds the current gain for the configurable Silence Hold (0.8 s by default) whenever the fast signal is absent or the lower threshold is crossed. This keeps rests and decaying notes from causing rapid target redistribution or gain chasing, while sustained inactivity still returns the channel toward unity. A manual row gesture opts that Child out. This is intentionally slow gain riding rather than transient compression, preserving instrument attacks and reducing audible pumping.

That original control law remains available as Live Ride. The optional Stable Level law is intended for mixed instrumental inputs whose event-to-event dynamics should remain intact. It derives the same weighted target from a configurable 2-30 second smoothed source-level estimate (8 seconds by default) rather than directly from each 400 ms reading, retains learned sources in the allocation while they rest, and holds changes inside the configurable Stable Deadband (1 dB by default). It fills one complete selected window before ordinary correction and caps that correction at the configurable Stable Speed (0.5 dB/s by default). Crest history is held for at least that same window and a recurring crest within 0.5 dB refreshes the hold, preventing gain recovery from following the spaces between transients. Only a real instantaneous ceiling violation bypasses the slow rate. A deterministic three-source drum/synth/guitar test compares cumulative trim movement against Live Ride and verifies that learned trims remain unchanged through a five-second rest. A separate 5.3-second sparse-crest loop regression verifies stable trim inside a 20-second window.

Evidence: SignalExchange/AUTO_BALANCE_RESEARCH.md.

A worked allocation example

For three active automatic sources with weights 0, 0 and +3 dB, the linear weights are approximately 1, 1 and 1.995, so W≈3.995 and 10*log10(W)≈6.015 dB. With a remaining target of −20 LUFS, their allocated targets are about −26.015, −26.015 and −23.015 LUFS. The featured source gets a larger share without raising the predicted total power budget.

With one source, its own weight cancels in b_i-10*log10(10^(b_i/10)), so changing that absolute weight does not raise the group's target. Manual rows consume measured post-gain power before automatic allocation. These are power-budget calculations; actual correlated audio summing, downstream effects and peak reconstruction still require measurement on the real output.

Measure the audio that actually reaches the output

Every Child applies the BS.1770 K-weighting response to its pre-gain stereo input and accumulates energy in an exact 400 ms sliding rectangular window. The resulting Momentary value is published to the Master. There is no loudness gate on this measurement; Activity Threshold is a product control gate used only to decide whether automatic gain should act.

Independently, the Master measures its post-limiter mono/stereo programme output. It maintains ungated 400 ms and 3 s windows, constructs Integrated gating blocks every 100 ms, and samples LRA at 10 Hz. Constant-memory 0.01 LU histograms retain counts and exact power sums, allowing every live Integrated update to recalculate the relative gate without audio-thread allocation. Pausing excludes audio from I/LRA and programme maxima while current M/S remain live. Reset clears I, LRA, maximum M/S, and maximum true peak simultaneously in either transport state.

This is an audio-path measurement, not a reconstruction from Child measurements. The shared bus carries pre-trim measurement summaries and control commands only; it does not carry sample audio and cannot represent unregistered tracks, downstream processing, summing, or final-bus gain. A conforming programme reading therefore requires Master to be inserted on the actual main output or measured group. Active Child rows alongside -inf EBU readings mean the control path is alive while the Master audio input is silent or unrouted.

The Master limiter sidechain reconstructs inter-sample peaks with a fixed 49-tap polyphase FIR:

• 4x below 96 kHz;
• 2x from 96 kHz to below 192 kHz;
• native sample rate at 192 kHz and above.

It uses a stereo-linked 5 ms lookahead and a final PCM sample clamp. This safety stage is designed to respect the selected production ceiling. It is not advertised as a certified delivery limiter.

Evidence: SignalExchange/CONFORMANCE.md.

What the independent fixtures established

When the private test fixtures are installed, CTest registers two optional runners.

The local archive used to establish this evidence was 91,631,421 bytes with SHA-256:

The official material is copyrighted by the EBU, used for technical testing only, excluded by Git, and not distributed with SignalExchange. The runners are registered only when the expected files exist under build/test-fixtures/ebu/files.

A second processor-level concert simulation runs eight real Children for 28 seconds with different spectra and input levels, changing three-/seven-/two-source ensembles, a 200 ms musical rest, a four-second rest, irregular 64–1024-sample blocks, a persistent ambient-noise channel, and a clipped stereo transient. Its current deterministic result settles at -19.56 LUFS, leaves the noise-only channel unboosted, preserves gain through the short rest, returns the long-resting channel to unity, and measures no more than the selected -6 dBFS output true-peak ceiling.

The host-lifecycle regression suspends and reactivates both processor types, proving that the Child's Momentary history and the Master's delayed audio cannot cross that boundary. A separate global allocator audit warms all lazy paths and then observes 2,000 Child-to-Master process() callback pairs with zero C++ heap allocations.

Evidence: SignalExchange/CONFORMANCE.md.

Applying the findings

Use slow, qualified boosts and a separate immediate attenuation path for a real ceiling violation. Test silence with both a slow loudness window and a fast activity indicator so the controller does not boost a decaying window after the musician has stopped. Hold crest history across sparse repeats to avoid recovering gain between transients.

Compare Live Ride and a learned Stable Level policy on the same multi-source schedule. Measure cumulative trim movement, rests, entrances and manual handoffs, not merely the final loudness. Then route the actual summed audio through the final meter and true-peak stage. IPC connectivity and a sample-peak ceiling cannot substitute for that final path.

Limits and open questions

The −20 LUFS working mix target is not the EBU −23 LUFS full-program delivery reference. The allocation law is product-specific, and applicable fixture passes do not constitute product certification or universal broadcast acceptance. Private EBU test audio is not redistributed. Host routing and native output measurements remain separate evidence from deterministic processor tests.

Research illustrations

Prismal Labs native editor capture, illustrating the corresponding software project. View full-size image

Primary references

• ITU-R BS.1770-5: loudness and true-peak measurement
• EBU Tech 3341: EBU Mode metering
• EBU R 128: loudness normalization

Research provenance

The methods and results above are drawn from these retained research records. Dates and product versions inside the article identify the scope of each experiment; publication review does not mean the hardware or numerical tests were rerun.

• SignalExchange/AUTO_BALANCE_RESEARCH.md — SHA-256 4357e8bc83791e7f74eb46adb62a2da769396eb99498ee53f85ec66fa7ad6dea
• SignalExchange/CONFORMANCE.md — SHA-256 a26114e7fa54bd1df1c776f94ee72710e70b940a71cb5b64e3d549aa014381e9

Related projects

SignalExchange Child · SignalExchange Master

All audio and hardware research · Audio Plugins