How Sonalix measures your mix
Last updated 6 min readTags:AnalysisSonalix
How Sonalix decodes uploads and calculates loudness, true peak, spectrum, stereo and tonal findings.
Sonalix decodes your upload, then measures the resulting samples. Levels and the main spectrum cover the file as a whole, subject to the loudness gates. The tonal comparison selects the main sections before it compares them with reference releases. For the order to read the results, see your mix report.
Audio input
FFmpeg decodes each file to 32-bit float at its native sample rate and channel layout. Float represents 24-bit PCM exactly and can hold values above full scale. Values from 32-bit integer and 64-bit float sources round to 24 significant bits, with relative error on the order of 150 dB below the sample for normal values.
Sonalix accepts mono and stereo. It classifies PCM, FLAC, ALAC, WavPack, APE, TTA, MLP and TrueHD codecs as lossless. This is a decoder classification, separate from the list of accepted upload formats. With a lossy upload, the decoded waveform determines the peaks and the top of the spectrum. Use a lossless export if you want those readings to reflect the original mix.
Loudness, sample peak and true peak
One FFmpeg process runs four branches. The ebur128 branch emits loudness and running peak metadata. A second branch pads the file with 0.5 s of silence to flush the true-peak resampler. A third counts samples and records each channel's sample peak. On stereo files, a fourth measures the difference between left and right to detect identical channels.
| Reading | Window or rule |
|---|---|
| Momentary loudness | 0.4 s window |
| Short-term loudness | 3.0 s window |
| Integrated loudness | Complete 0.4 s blocks, absolute gate of −70 LUFS, then a relative gate 10 LU below the loudness calculated after the absolute gate |
| Loudness range (LRA) | Complete 3.0 s windows with no padding, followed by gates at −70 LUFS and −20 LU |
The relative gate for integrated loudness follows ITU-R BS.1770. EBU R 128 specifies the broadcast recommendation, with LRA defined separately in EBU Tech 3342.
The level engine evaluates momentary and short-term maxima every 100 ms. Its momentary maximum can under-read by up to 0.58 LU compared with a finer time grid. Digital silence has no reportable integrated loudness, LRA or peak. An upload with no block above −70 LUFS is rejected. Integrated loudness needs at least one complete 0.4 s block. LRA needs at least five gated 3 s windows and is provisional below 60 s. The minimum prevents a percentile artefact seen in FFmpeg, where a constant 3.0 s tone could otherwise yield 20 LU.
True-peak validation
ebur128 resamples to 192 kHz for true peak. Sonalix reports the greater of that reading and the sample peak.
| Sample rate | True-peak status |
|---|---|
| 44.1 and 48 kHz | Validated against EBU Tech 3341 cases 15–19 |
| 32, 88.2, 96 and 176.4 kHz | Provisional |
| Other rates | Unavailable |
At 192 kHz and above the engine does not oversample. Validated readings use 0.1 dB increments, with errors on the test signals between −0.4 and +0.2 dB. If the highest peak occurs in the first 100 ms, resampler start-up can add an over-read of about 0.5 dB. These limits matter near full scale. See metric thresholds.
PLR for mono and stereo
Peak-to-loudness ratio (PLR) is true peak minus integrated loudness. A uniform gain change shifts both by the same amount when the same material clears the gates, subject to display resolution. For a one-channel file, Sonalix uses its dual-mono loudness for PLR. This adds 3.01 LU to the one-channel reading and gives the same PLR as a two-channel file with that same signal in both channels at unity. Crest factor is outside the report's measurements.
Spectrum and stereo
Sonalix uses a periodic Hann window of N samples with a hop of N/4. N is the smallest power of two that gives a bin width of at most 1.465 Hz. At 44.1 and 48 kHz it is 32,768 samples. Zero extension at both ends gives every sample four frames of coverage. Scaling preserves energy between the time and frequency domains. The whole-track spectrum averages linear mean-square power per bin, including silence, before conversion to decibels.
| Measurement band | Range |
|---|---|
| Sub | 20–60 Hz |
| Bass | 60–120 Hz |
| Low-mid | 120–350 Hz |
| Mid | 350–2,000 Hz |
| Upper-mid | 2,000–5,000 Hz |
| Presence | 5,000–10,000 Hz |
| Air | 10,000–20,000 Hz |
For stereo, M = (L + R) / 2 and S = (L − R) / 2. Correlation is (P_M − P_S) / √(P_L × P_R), where P is mean-square power. It runs from −1 to +1, at zero lag and without mean removal. The mono-sum change compares the power of the mono fold-down with stereo power in dB and is at most 0 dB. The correlation time series uses 3 s windows. Low-band stereo checks cover 20–120 Hz.
Tonal comparison
Sonalix selects active seconds within 4 dB of the loudest 20 s. If fewer than 8 s qualify, it uses all active seconds. It compares each selected second with the reference curve and removes its overall level using the median difference from 100 Hz to 8 kHz. It then takes the median of each region's per-second mean difference. A brief peak therefore has less effect than a persistent one.
| Tonal comparison region | Range |
|---|---|
| Sub-bass | 20–45 Hz |
| Bass | 45–120 Hz |
| Low mids | 120–400 Hz |
| Mids | 400–1,000 Hz |
| Upper mids | 1,000–3,500 Hz |
| Presence | 3,500–8,000 Hz |
| Top end | 8,000–16,000 Hz, with comparison limited to 14 kHz |
A build-up spans about half an octave between 30 Hz and 8 kHz. A narrow resonance is a one-sixth-octave point above its neighbours between 120 Hz and 6 kHz. The report shows up to two of each.
Reference balances
The reference balance is the median of commercial releases measured over their main sections using the same method. The general balance pools 33 releases. Pop has 6, hip-hop 5, electronic 6, house 8 and rock 8. A genre with one to four references uses the general curve moved halfway towards its examples. With none, the report uses a declared fallback. It names the balance it used and caps tonal findings at medium for sparsely represented genres and fallbacks.
The reference files are lossy streams. Sonalix excludes content above 14 kHz from the comparison. The calibration is based on the measurements, without a listening panel. The reference releases calibrate tonal comparison. Integrated loudness, true peak, PLR and LRA describe the uploaded file without a commercial target.
From a measurement to a finding
Rules use the measurements to decide when to flag a condition. Their status and the effect of mix stage are listed in metric thresholds and target ranges. If another meter gives a different result, see why your DAW meter and Sonalix disagree. The supported file formats guide covers how the export can affect the result.