Load any convolution reverb and you are loading a recording of a space. But that recording is not a mic hanging in a cathedral while someone fires a starter pistol. The old impulsive excitation methods, clap guns, popped balloons, blanks, produce signals too short and too quiet for accurate measurements in anything but an anechoic chamber. The algorithm that replaced them was published by Angelo Farina at AES Convention 108 in Paris in 2000: "Simultaneous Measurement of Impulse Response and Distortion with a Swept-Sine Technique." It works by playing a tone that glides from 20 Hz to 20 kHz through a loudspeaker and recording what comes back.
What an Impulse Response Actually Is
An impulse response (IR) is the acoustic fingerprint of a linear time-invariant (LTI) system. Play a theoretical impulse, a signal that is infinitely brief with flat frequency content, and whatever comes out is the IR. The room's modes, reflections, absorption, diffusion and decay are all encoded in that single time-domain waveform. Convolve any dry audio with the IR and you hear that dry signal played in the same room.
The problem is that a real impulse has almost no energy. A pistol shot has perhaps 1-2 ms of useful signal across 20 Hz to 20 kHz. In a 1-second reverb tail you might have 50 dB of dynamic range to measure. Pistol shots in real rooms deliver maybe 30 dB above the noise floor at best. The reverberation tail gets buried.
The solution is to use a longer signal that can be inverted back into an impulse after the fact.
What MLS Got Wrong
Before Farina, the dominant method was Maximum Length Sequence (MLS). An MLS is a binary pseudo-random sequence with a spectrum that is flat to within a small ripple, white noise in disguise. You play the sequence through a speaker, record the output, cross-correlate with the known sequence, and extract the IR. With enough averaging you can push the noise floor down by 20-30 dB.
MLS works until any of three things happen: the speaker or system under test is nonlinear, the room temperature drifts slightly between measurements, or someone opens a door. The cross-correlation operation assumes strict time invariance. A one-sample drift in a 20-second MLS spread causes an artefact equivalent to a second reflection appearing in the IR at a random location. Nonlinearity in the loudspeaker, which is unavoidable at any useful playback level, folds harmonic distortion back onto the linear IR instead of isolating it. You cannot tell whether a measured reflection is real or an artefact of the woofer's voice coil hitting its limits at 80 Hz.
The Exponential Sine Sweep
The ESS signal is a sine wave whose instantaneous frequency increases exponentially with time:
s(t) = sin[ K ( e^(t/L) - 1 ) ]where K = ω₁ · L and L = T / ln(ω₂/ω₁). Here ω₁ and ω₂ are the start and end angular frequencies, and T is the total sweep duration. The frequency at time t is:
f(t) = f₁ · (f₂/f₁)^(t/T)Exponential means each octave gets the same amount of time. The sweep spends as long on the 80-160 Hz range as on the 1280-2560 Hz range. Because human hearing perceives frequency on a logarithmic scale, and because the low end contains most of the acoustic energy in a room and most of the nonlinearity in loudspeakers, this is exactly where you want measurement time concentrated.
The resulting power spectrum rolls off at 3 dB per octave, the same slope as pink noise. Playing a flat sweep through a speaker that has more energy at low frequencies than high would overdrive the woofer and leave the tweeter contribution in the noise. The ESS signal's pink spectrum matches a typical loudspeaker's acoustic output profile far better than white-spectrum MLS.
Deconvolution: Undoing the Sweep
After recording the room's response to the sweep, you extract the IR by convolving the recording with a time-reversed, frequency-weighted version of the original sweep signal. This inverse filter is pre-computed from the known sweep parameters and is not sensitive to loudspeaker output level, which means you do not need to calibrate the playback chain precisely.
In the frequency domain the operation is simple: divide the spectrum of the recorded response by the spectrum of the sweep. This is equivalent to subtracting the phase and spectral shape of the sweep from whatever came back, leaving only what the room added.
Harmonic Distortion Falls Out for Free
Here is the part that makes the ESS method qualitatively different from everything before it. Because the sweep is exponential, the nth harmonic of any frequency in the sweep is always reached by the fundamental sweep at a later time in the signal. When you play 200 Hz, the 400 Hz component of the sweep has already passed. When you play 400 Hz, the 800 Hz component has already passed.
After deconvolution, anything that was generated as harmonic distortion, by the loudspeaker, the amplifier, or the room itself, arrives at the analysis as a separate component displaced earlier in time than the linear IR. The 2nd harmonic distortion products collapse into a narrow spike that appears a fixed number of samples before time zero. The 3rd harmonic appears even earlier. Each harmonic order occupies a distinct time slot in the deconvolved output.
This means you can window the deconvolved response to isolate just the linear IR (everything after a gate positioned just before time zero), or you can look at the pre-zero region to read off the harmonic distortion spectrum of the speaker you used for measurement. A compression driver running out of headroom shows up as a visible 2nd-harmonic spike. An amplifier with crossover distortion shows up as a 3rd-harmonic spike. These artefacts are physically separated from the acoustic measurement you care about, and you can discard them cleanly.
MLS cannot do this. Loudspeaker distortion in MLS is cross-correlated back into the main IR as a smear of false reflections distributed across the entire measurement window.
Practical Measurement Workflow
The standard workflow using a tool like Room EQ Wizard:
- Place a calibrated measurement microphone at the listening position.
- Route the sweep through your DAW or measurement software to your room speakers at a level that is loud enough for adequate SNR but not loud enough to drive the speakers into hard clipping.
- Record the output on the microphone input. Typical sweep durations are 8 to 30 seconds; longer sweeps give better low-frequency SNR at the cost of requiring a stable acoustic environment throughout.
- The software deconvolves the recording and returns an IR that can be exported as a WAV file.
- Optionally average two or more sweeps to push the noise floor down further.
The result is not just a reverb impulse. The same measurement gives you the frequency response at the listening position, the reverberation time across frequency bands (the Schroeder plot), the early decay time, and a waterfall graph showing which frequencies ring the longest. Every piece of data needed to understand a room comes from a single sweep measurement.
Using ESS for Non-Room Captures
The method is not limited to rooms. Sound designers and engineers routinely use ESS to capture:
- Hardware reverb units: spring tanks, plate reverbs, vintage convolution hardware
- Amps and cabinets: the loudspeaker impulse response for use as a guitar cab IR
- Acoustic instruments: the body resonance of a guitar or drum shell for physical modeling
- Small spaces: stairwells, pipes, resonant objects, concrete tunnels
In every case you are extracting a linear transfer function while simultaneously diagnosing nonlinear behaviour in the device under test. The spring reverb in your rack has a frequency-dependent distortion character from its spring coils. After ESS deconvolution you can see exactly how bad it is at each frequency.
SNR Versus MLS
Comparative studies have found that ESS measurements deliver roughly 15 dB more usable dynamic range than MLS in real-world conditions with loudspeakers operating at normal studio levels. In very quiet rooms with perfect loudspeakers and no time variance, the methods converge. In the real world, which involves nonlinear loudspeakers, HVAC noise, and doors that occasionally close, the exponential sweep wins cleanly.
The summary: MLS is theoretically elegant but practically fragile. ESS is theoretically slightly more complex but practically robust. When someone hands you a convolution reverb impulse response of a famous concert hall, there is a very good chance it was measured with a sweep and extracted by a modified version of Farina's 2000 deconvolution procedure.
You can replicate the entire workflow with a measurement microphone, Room EQ Wizard (free), and a couple of hours. The most famous spaces in audio are just sweep recordings with the sweep subtracted out.