Louder, tighter, phase-safe: what changed in the Tensormix engine

A measurement-driven follow-up to our original mastering comparison.

When we first benchmarked Tensormix against the other automated services, our engine made a deliberate choice: target moderate loudness, preserve dynamic range, and lean into stereo width. On the same three tracks we used then, Tensormix landed around -14.5 to -15 LUFS - comfortably below LANDR and eMastered, but above SoundCloud.

We've chosen to adjust our approach. Two things drove this - industry trends, including how the majority of today's commercial music is mastered, and what user's actually prefer.

The 2026 Grammy Record of the Year nominees averaged -8.41 LUFS - roughly 5.6 dB louder than Spotify's -14 LUFS reference - and half of them ran their true peaks over 0 dBTP. Streaming normalization is a floor commercial releases routinely ignore, not a ceiling.

So we've made some changes to the loudness, dynamics, and stereo handling of the engine. This post re-runs the exact same benchmark — same source mixes, same competitors, same analyzer - so the changes are visible in the numbers.

The test

Three source mixes, spanning the electronic styles most of our users produce:

• Drown — liquid drum & bass
• I Want It — deep house / melodic techno
• Phosphorescence — melodic breakbeat

Each mix was mastered through the production engine on its adaptive default (no genre override), then analyzed alongside masters of the same mix from eMastered, LANDR, and SoundCloud, plus the unmastered source as a reference.

A quick key to the columns:

• LUFS — integrated loudness (BS.1770 / EBU R128).
• True Peak — inter-sample peak in dBTP. Above 0.00 risks clipping on conversion.
• LRA — loudness range, in LU. Our macro-dynamics measure.
• Crest — peak-to-RMS, in dB. Short-term punch.
• PLR — peak-to-loudness ratio (True Peak − LUFS).
• SpecCtr — spectral centroid in Hz; higher = brighter.
• Low / LowMid / Mid / High / Tilt — energy per octave band; Tilt = High − Mid.
• Sub Corr — L/R correlation below 120 Hz. Near 1.00 means the bass is mono and won't collapse on a large system.
• Hi Corr — L/R correlation above 120 Hz. Lower = wider stereo image.

Loudness

Track	Then (original post)	Now
Phosphorescence	-14.52 LUFS	-10.40 LUFS
I Want It	-15.02 LUFS	-10.72 LUFS
Drown	—	-10.54 LUFS

That's roughly +4 dB across the board. The engine now lands in the same competitive band as LANDR (-9.72 to -10.89 across these tracks) and eMastered (-10.31 to -11.07), rather than sitting a full genre quieter. We now sit squarely in the range that commercial masters occupy.

True peak

Getting louder usually means flirting with the 0 dBTP line - which we didn't want to do, so we made sure that our limiting would handle stray inter-sample peaks.

Side note: our loudness handling is tested against the official EBU spec, same as our plugins (more on plugins in the coming weeks).

Track	TENSORMIX	eMastered	LANDR	SoundCloud
Drown	-0.10	+0.74	-0.25	+0.48
I Want It	-0.10	-0.07	-0.26	-0.04
Phosphorescence	-0.10	+0.29	-0.10	+0.36

Tensormix holds a clean -0.10 dBTP ceiling on every track. eMastered and SoundCloud both push true peaks above 0 dBTP on the bass-heavy material — the exact condition that produces clipping after lossy encoding. We hit competitive loudness and keep the ceiling intact at the same time.

Dynamics

The biggest internal change was how we handle macro-dynamics. There is no such thing as a universal "correct" dynamic range - a sparse breakbeat and a wall-of-sound techno drop should not be flattened to the same LRA just to match some sort of genre-specific average. This is where human mastering makes the biggest difference - it's cruicial that you adapt your dynamics strategy to your source, and don't try to apply a preset or chase a specific target.

Track	Variant	LUFS	LRA	Crest	PLR
Drown	SOURCE	-18.01	9.77	18.82	17.70
	TENSORMIX	-10.54	5.38	12.84	10.44
	eMastered	-11.07	7.12	13.60	11.81
	LANDR	-9.72	6.88	9.70	9.47
	SoundCloud	-13.57	8.32	15.13	14.06
I Want It	SOURCE	-16.45	8.45	16.52	14.46
	TENSORMIX	-10.72	5.38	13.94	10.62
	eMastered	-10.31	6.48	12.02	10.25
	LANDR	-10.71	7.74	12.47	10.45
	SoundCloud	-13.96	7.55	15.83	13.92
Phosphorescence	SOURCE	-13.42	11.03	14.42	13.78
	TENSORMIX	-10.40	5.93	12.20	10.30
	eMastered	-10.66	8.16	11.92	10.95
	LANDR	-10.89	9.92	11.65	10.79
	SoundCloud	-13.74	9.67	15.17	14.10

At this loudness, LRA naturally tightens and crest factor becomes an interesting metric to keep an eye on. On Drown, Tensormix retains 12.84 dB of crest against comparable loudness, and on I Want It we hold 13.94 dB, the most short-term punch of any master in the group.

In practice this means we're achieving our target loudness without sanding off the transients.

Phase and correlation

A wide, exciting stereo image is worthless if it folds in on itself the moment the track plays on a club PA, a festival rig, or any system that sums to mono. The low end is the most dangerous: out-of-phase bass doesn't just narrow, it disappears on a mono subwoofer.

So we treat the two halves of the spectrum differently. Below 120 Hz the engine keeps the image tight to mono; above it, it widens. The benchmark shows both behaviors at once:

So given how much of our user base makes electronic music, we pay a lot of attention to low-end frequencies and mono-compatability:

Track	Variant	Sub Corr (higher = mono-safe bass)	Hi Corr (lower = wider)
Drown	SOURCE	0.84	0.60
	TENSORMIX	0.95	0.51
	eMastered	0.80	0.56
	LANDR	0.92	0.57
	SoundCloud	0.83	0.56
I Want It	SOURCE	0.98	0.63
	TENSORMIX	0.99	0.44
	eMastered	0.98	0.59
	LANDR	0.98	0.61
	SoundCloud	0.98	0.61
Phosphorescence	SOURCE	0.87	0.51
	TENSORMIX	0.98	0.45
	eMastered	0.85	0.50
	LANDR	0.92	0.51
	SoundCloud	0.87	0.49

A footnote, beyond electronic

The benchmark above is built around electronic material because that's most of what our users make. But the dynamics, true-peak, and phase work isn't genre-specific. Perhaps one of the biggest winners of this whole upgrade is Rock and Metal:

Metric	Source	TENSORMIX
LUFS	-14.74	-10.31
True Peak	-1.79	-0.10
LRA	4.66	3.42
Crest	17.11	13.25
Sub Corr	0.88	0.97
Hi Corr	0.32	0.41
Tilt	-2.93	+1.27

The same signatures show up: +4.4 dB of loudness to a competitive -10.31 LUFS, the true peak held to a clean -0.10 dBTP, the sub end tightened from 0.88 to a near-mono 0.97, and the tonal balance opened from a dark -2.93 tilt up to a present +1.27 - exactly the top-end clarity a dense guitar mix needs. The engine carries the same discipline across genres.

And while we don't have any data to back this up, metal simply sounds huge now.

Where this leaves us

The thesis has shifted since the last post, and the measurements show it. Our goal is to continue to bridge the gap between existing algorithmic mastering services and human engineers, and these upgrades are a step towards Tensormix being a competitive commercial mastering service.