Method · every number on this site
How we measure.
A number without a method is decoration. This is the method behind the deltas on every product page and the chain on the homepage, so you can judge them — or reproduce them.
The rules
Same method, every unit.
If a product page shows a figure, it came from this process. If this process can't produce a figure for a claim, the claim isn't made.
MaterialReal program material — vocals, drums and 808s, full mixes — chosen per unit for the job it does. The homepage chain runs one melodic-rap vocal through five units in order; every product demo is a single unit on its own program.
One instanceEach demo is exactly one instance of the plugin at the settings listed under the player. No hidden chain, no post-processing, no mastering on the A or the B.
Level matchingThe processed take is gain-matched to the dry take on K-weighted integrated loudness (ITU-R BS.1770). The match is a pure gain change applied after the plugin; nothing else touches the audio. Loudness deltas shown on limiters are the exception — that gain is the product.
LatencyEvery stage is time-aligned to the dry take by the latency the plugin reports, verified by cross-correlation, so before/after comparisons are sample-accurate.
What Δ crest meansCrest factor is peak minus RMS. A negative Δ crest means the unit reduced peaks relative to the body of the signal — what compression, limiting and saturation do. A positive Δ crest means transients came forward.
What Δ sibilance and Δ body meanBand energy deltas: sibilance is 5–10 kHz, body is 120–400 Hz, measured on the level-matched pair. They show where in the spectrum the unit did its work.
True peakMeasured with a 4× oversampled true-peak meter on the delivered file. Limiter demos state the ceiling; the true-peak figure shows it held.
FormatDemos are delivered as AAC at about 256 kb/s, 48 kHz. The generic deltas (Δ loudness, Δ crest, band deltas, true peak) are taken on the lossless renders before encoding. The job-specific figures below are taken on the delivered AAC files — what you actually hear — so encoder overshoot shows up in them rather than being hidden.
Bench receiptsBefore a build ships it passes a battery per unit: null tests against the design reference, aliasing floor, meter accuracy against measured gain reduction, program-dependent behaviour, and host validation across AU, VST3 and AAX. The receipts printed on a product page are the checks that build passed.
Stem separationWhere a chain demo needs an isolated vocal, the vocal is separated with a Demucs-family model; on the listening path the untouched original mix is used and only the vocal's processing delta is layered on, so the beat never passes through separation.
What isn't measuredInstrument demos (FUSION, CADENCE) compare a stock synth and the instrument playing the same MIDI; there is no processed input, so they carry no before/after figure. A SCOUT demo is a detector readout, so its only figure is that the audio is untouched. Where a number doesn't exist, the page doesn't invent one.
Measured for the job
Each unit, on its own terms.
Loudness and crest don't tell you whether a pitch corrector corrected pitch. Each product page leads with one to three figures chosen for what that unit does, computed by one script from the two demo files you can play. Where the after measures worse on its own job, or doesn't move, the page says so — the numbers aren't filtered.
Pitch error
HALOYIN pitch tracking (de Cheveigné & Kawahara) at 16 kHz, 64 ms frames, confidence threshold 0.12, 100–900 Hz. Each confidently voiced frame's distance to the nearest equal-tempered semitone (A = 440 Hz), in cents; we publish the median and the share within ±10 ¢. Limits: it runs on the full mix, so only the frames where the voice dominates the beat are counted (under 10 % of the clip), and vibrato and slides count as error on both sides.
HALOYIN pitch tracking (de Cheveigné & Kawahara) at 16 kHz, 64 ms frames, confidence threshold 0.12, 100–900 Hz. Each confidently voiced frame's distance to the nearest equal-tempered semitone (A = 440 Hz), in cents; we publish the median and the share within ±10 ¢. Limits: it runs on the full mix, so only the frames where the voice dominates the beat are counted (under 10 % of the clip), and vibrato and slides count as error on both sides.
Word-to-word spread / Loudness range
EA-2A, EA-3A, EA-76, POCKET, LARYNX, GLUEK-weighted (BS.1770) loudness in 400 ms windows, gaps more than 20 LU below integrated removed, P90 − P10. Loudness range is the EBU Tech 3342 method (3 s windows, P95 − P10), used on the longer clips; it matches ffmpeg's ebur128 to 0.1 LU. Limits: on vocal demos it is measured on the whole mix, beat included, so the beat dilutes the vocal's own leveling.
EA-2A, EA-3A, EA-76, POCKET, LARYNX, GLUEK-weighted (BS.1770) loudness in 400 ms windows, gaps more than 20 LU below integrated removed, P90 − P10. Loudness range is the EBU Tech 3342 method (3 s windows, P95 − P10), used on the longer clips; it matches ffmpeg's ebur128 to 0.1 LU. Limits: on vocal demos it is measured on the whole mix, beat included, so the beat dilutes the vocal's own leveling.
Hit-to-hit spread
GRIP, EA-60Peaks of a 2 ms envelope at least 90 ms apart and within 24 dB of the loudest hit; we publish their standard deviation in dB. Limits: a crude onset picker — ghost notes and flams can merge or split hits.
GRIP, EA-60Peaks of a 2 ms envelope at least 90 ms apart and within 24 dB of the loudest hit; we publish their standard deviation in dB. Limits: a crude onset picker — ghost notes and flams can merge or split hits.
Ess peaks
EA-DS, LARYNX5–10 kHz energy relative to broadband energy in 10 ms frames, quiet frames dropped; the 95th percentile is the esses. The median of the same ratio (between the esses) shows whether the cut is dynamic or just a static EQ.
EA-DS, LARYNX5–10 kHz energy relative to broadband energy in 10 ms frames, quiet frames dropped; the 95th percentile is the esses. The median of the same ratio (between the esses) shows whether the cut is dynamic or just a static EQ.
Integrated loudness, true peak, overs
CREST, EA-L, APEXBS.1770 integrated loudness (matches ffmpeg's ebur128), true peak by 4× oversampling, and the count of oversampled samples above the ceiling. Measured on the delivered AAC, so any encoder overshoot is counted against the unit.
CREST, EA-L, APEXBS.1770 integrated loudness (matches ffmpeg's ebur128), true peak by 4× oversampling, and the count of oversampled samples above the ceiling. Measured on the delivered AAC, so any encoder overshoot is counted against the unit.
Measured EQ gain
EA-EQ, EA-PQ, HERITAGEThe linear transfer function estimated from the pair (cross-spectrum over input spectrum, 16384-point Welch, latency removed), averaged over a ⅓ octave at the target frequency and expressed relative to the 1 kHz ⅓ octave so the level match cancels out. On EA-EQ, a static EQ, it reads −3.9 dB where −4 dB was set.
EA-EQ, EA-PQ, HERITAGEThe linear transfer function estimated from the pair (cross-spectrum over input spectrum, 16384-point Welch, latency removed), averaged over a ⅓ octave at the target frequency and expressed relative to the 1 kHz ⅓ octave so the level match cancels out. On EA-EQ, a static EQ, it reads −3.9 dB where −4 dB was set.
Mud peaks
CONTOURThe frequency of the deepest average cut between 100 Hz and 2 kHz is found from the same transfer estimate; then the change in the 95th percentile (peaks) and the median of that ⅓ octave's 50 ms level relative to 1 kHz. A dynamic EQ should move the peaks more than the median.
CONTOURThe frequency of the deepest average cut between 100 Hz and 2 kHz is found from the same transfer estimate; then the change in the 95th percentile (peaks) and the median of that ⅓ octave's 50 ms level relative to 1 kHz. A dynamic EQ should move the peaks more than the median.
Phone-band 808 / Added harmonics
KNOCK, EA-SAT, ENRICH, HERITAGE, MOLTPhone-band 808 is 200 Hz–1 kHz energy over 30–100 Hz energy. Added harmonics is the share of the after's energy that is not linearly predictable from the before (energy-weighted 1 − coherence, latency removed); a static EQ reads about −27 dB, the floor set by the codec and the estimator. Limits: any time-varying gain also raises it, so it is a proxy for added harmonics, not a THD figure.
KNOCK, EA-SAT, ENRICH, HERITAGE, MOLTPhone-band 808 is 200 Hz–1 kHz energy over 30–100 Hz energy. Added harmonics is the share of the after's energy that is not linearly predictable from the before (energy-weighted 1 − coherence, latency removed); a static EQ reads about −27 dB, the floor set by the codec and the estimator. Limits: any time-varying gain also raises it, so it is a proxy for added harmonics, not a THD figure.
Air band
ENRICH10–16 kHz energy as a share of full-band energy. The before was low-passed at 10 kHz, so energy above it has to come from the unit.
ENRICH10–16 kHz energy as a share of full-band energy. The before was low-passed at 10 kHz, so energy above it has to come from the unit.
Added decay (RT60)
AETHER, EA-VERBAn impulse response estimated from the pair (cross-spectral, 500 Hz–4 kHz, where the voice sits above the beat), then a straight line fit to its energy decay from the wet peak down to 5 dB above the estimate's own noise floor, extrapolated to −60 dB. Needs 15 dB of clean decay or it isn't published. Limits: the beat under the vocal biases the wet level, so we publish decay time, not wet level; the fit covers far less than 60 dB.
AETHER, EA-VERBAn impulse response estimated from the pair (cross-spectral, 500 Hz–4 kHz, where the voice sits above the beat), then a straight line fit to its energy decay from the wet peak down to 5 dB above the estimate's own noise floor, extrapolated to −60 dB. Needs 15 dB of clean decay or it isn't published. Limits: the beat under the vocal biases the wet level, so we publish decay time, not wet level; the fit covers far less than 60 dB.
Echo time, repeat decay, echo pan
EA-ECHO, DRIFTEcho time is the strongest peak of the same impulse-response estimate between 40 ms and 1 s (5 ms bins); on EA-ECHO it reads 318 ms against 317 ms set. Repeat decay is the second repeat relative to the first. Echo pan is the left-minus-right level of each repeat — a sign flip would be ping-pong.
EA-ECHO, DRIFTEcho time is the strongest peak of the same impulse-response estimate between 40 ms and 1 s (5 ms bins); on EA-ECHO it reads 318 ms against 317 ms set. Repeat decay is the second repeat relative to the first. Echo pan is the left-minus-right level of each repeat — a sign flip would be ping-pong.
Stereo width, correlation, pan movement
EXPANSE, MOTION, VOICES, reverbsWidth is side/mid energy in dB (more negative is narrower), per band where the claim is per band. Correlation is the L/R Pearson coefficient. Pan movement is the standard deviation of L/R balance over 50 ms frames. Measured on the whole mix.
EXPANSE, MOTION, VOICES, reverbsWidth is side/mid energy in dB (more negative is narrower), per band where the claim is per band. Correlation is the L/R Pearson coefficient. Pan movement is the standard deviation of L/R balance over 50 ms frames. Measured on the whole mix.
Samples changed
SCOUTThe count of decoded samples that differ between the two sides. SCOUT is analysis-only, so the right answer is zero.
SCOUTThe count of decoded samples that differ between the two sides. SCOUT is analysis-only, so the right answer is zero.
Hear it applied