How Room Correction Software Works

Strip the interfaces away and every correction product does the same five things. It plays a known test signal through the system and records what arrives at a microphone. It computes the system's response from the two, magnitude against frequency and usually phase and timing as well. It compares that response to a target and generates filters to close the gap, within limits. It sets delays and gains so every source arrives together at matched level. And it measures again with everything active to confirm the result. Products differ in how honestly each stage is done, and the first stage constrains all the others, because filters generated from bad data are precise about the wrong thing.

The measurement stage needs a signal with energy at every frequency and a way to compare what was sent with what arrived. Three approaches dominate. The swept sine plays one frequency at a time from low to high, which concentrates all the system's output into a narrow band at each moment, giving excellent signal-to-noise ratio and a clean impulse response. Sweeps also push harmonic distortion out of the measurement window, so the result reflects the linear response alone. The cost is that a sweep is a snapshot: one position, one moment, and the microphone and room must hold still for it. Background noise during the sweep lands directly in the data, and anything time-varying in the chain smears it.

MLS, the maximum length sequence, is a pseudorandom noise signal whose correlation properties let software recover the impulse response from a continuous broadband playback. It was the standard before fast sweep processing arrived, and it still appears in older tools. Its weaknesses are known: distortion anywhere in the chain folds back into the measurement as noise rather than being separated out, and it is more sensitive than the sweep to anything shifting during the capture. For most purposes the swept sine has replaced it.

The third approach is a live transfer function: play continuous pink noise, capture it, and compute the response as a running comparison between what is being sent and what the microphone hears, averaged over time. Alongside the response, the software computes coherence, a measure per frequency of how consistently the microphone signal follows the test signal. Coherence turns the measurement's honesty inward: bands corrupted by background noise, reverberant confusion or nulls show low coherence, and the software can weight or exclude them instead of correcting fiction. Because the measurement runs continuously, it also survives the real world differently. Noise averages down rather than embedding itself, the display responds while you move a speaker or a subwoofer, and moving the microphone through the listening area builds a spatially averaged response rather than a single point. This is the method Omnissiah uses, and the real-time measurement guide covers it in depth.

The sweep remains the right tool when you want a pristine impulse response of one fixed configuration, and it is what most products use. The live transfer function trades single-shot precision for statistical honesty and a measurement that reflects an area and a duration rather than a point and an instant, which is closer to how the system is actually heard.

The raw response is not corrected as-is. Software smooths it, typically to fractions of an octave, because the ear does not hear the microphone's every wiggle and correcting them produces filters that are wrong a centimetre away. It restricts the correction range, since above the room's transition region the measured response at one point stops representing what reaches the listener, and correction there does more harm than good. And it compares the result to a target curve, which is deliberately not flat, a topic the target curves guide covers on its own.

Filter generation is then a fitting problem: choose filters that move the smoothed response toward the target, subject to limits. The limits carry most of the engineering judgement. Boosts are capped because a boost spends amplifier headroom and driver excursion, and a boost into a null spends both for nothing. Narrow corrections are treated more cautiously than broad ones. Low-coherence data is excluded rather than corrected. Products doing this with minimum-phase IIR filters fix magnitude and phase of minimum-phase anomalies together, and the choosing guide explains why that matches the physics of what correction can validly fix. Alignment completes the set: inter-source delays and gains, ideally computed from the corrected responses so the filters' own phase behaviour is included.

The last stage is the one most often skipped and the only one that proves anything: measure again with the correction running. Prediction and reality diverge for good reasons, filters interact with each other, hardware imposes its own resolution, and the room does not care what the software expected. A product that re-measures and shows you the corrected response closes the loop. If yours does not, close it yourself with REW and a measurement microphone, and listen to well-known material before deciding the correction earned its place.