“Trust your gut before your head gets in the way”

Josephus Miller

Overview

Tensormix was released about 6 months ago.

I’ve wanted to do a technical writeup for it since it first came out, but for one reason or another I’ve put it off until now. So, this is a long overdue technical breakdown of what Tensormix is, how it performs, and how it stacks up against other offerings on the market.

This isn’t meant to be an attempt to sell you on the product, or to claim that it’s better than [insert competitor name here] - it’s simply meant to serve as a technical breakdown of a nerdy idea that I’m pretty proud of.

The idea

I’ve been making music in some form since I was about 12.

Music has always been a constant in my life. After seeing some minor success in it towards the tail end of my teenage years, I spent longer than I care to admit attempting to turn it into a career. Eventually I figured out that I was a far better programmer than a musician, but music has remained a hobby that I still dabble in on occasion.

At some point in the last couple of years I found myself talking to ChatGPT about something audio engineering related. Being able to provide screenshots of monitoring plugins and spectrum analysers meant that the advice that it gave me was directly applicable to whatever I was working on at the time.

If you’ve been making music for any reasonable length of time you probably know that when it comes to audio engineering, a huge chunk of advice that you hear on the internet (especially those low-effort Instagram infographics) is cookie cutter templating that might apply in some situations, but is by no means something that you can repeatedly apply to get the same results.

Music is unique. No 2 kicks are the same, no two vocalists can be EQued using the same settings, and those instrument frequency charts that tell you that you must high pass everything above a specific frequency for instrument X are ridiculous.

So talking to ChatGPT about audio when it understood the context, like the fundamental frequencies of the drums I was working with, felt very powerful. And the advice that it gave was spot on - afterall, this thing is trained on every textbook on audio engineering in existence (as well as the aforementioned Instagram infographics of course).

And this is where the idea for Tensormix came from - LLMs know audio engineering, and I’m sure someone at OpenAI would probably say that ChatGPT is the equvalent of a PhD in Audio Engineering. So what would happen if we let it drive the bus? Give it access to a bunch of tools - EQ, compressor, limiter etc - and let it run wild?

This sounded really cool, and the nerd in me wanted to build this immediately. The scope of this was enormous, and my DSP knowledge was severely lacking - so naturally, I decided to build an audio engineering agent.

Additionally, I’m not a very big fan of the direction that the music industry is heading in, with how lovingly record labels seem to have embraced generative AI - so a part of me wanted to try to make use of LLMs to build an alternative that didn’t attempt to prompt real musicians out of existence.

The results

Ok so I know this isn’t the most “scientific” testing methodology, but the basic idea for this test was to take three tracks, run them through the major mastering services, and measure their output in terms of key metrics, like dynamics, loudness, tonal balance etc.

(In actual dev testing and engine benchmarks there are A LOT more than 3 tracks)

The 3 tracks that we selected for testing all fit different scenarios:

Track 1 - Phosphorescence: I’d describe this as melodic breakbeat. It’s light and airy, high-end heavy, and most importantly for this test - it’s already quite hot to begin with, and goes in with very little headroom.

Track 2 - I Want It 120BPM deep house track with elements of techno thrown in. Sub heavy, not too aggressive, and fairly chill. And unlike the first track, this has a tonne of headroom for Tensormix to go crazy with limiting and gain staging.

Track 3 - Drown Liquid D&B, probably fittinng more tightly into a specific genre than the other 2 tracks. Plenty of headroom, very OK mix with the drums lacking punch and power.

All of the test tracks and their respective masters can be found on Arktide here

Track 1 - Phosphorescence

Track 2 - I Want It

Track 3 - Drown

Tensormix vs LANDR vs eMastered vs SoundCloud

Now, to reduce bias as much as possible, I’ve pasted these results into ChatGPT and asked it to provide an objective analysis - here’s what it had to say:

Link to chat

(ChatGPT’s) Objective Analysis of Online Mastering Results

This analysis compares the original mixes against automated masters from TENSORMIX, LANDR, eMastered, and SoundCloud across three tracks. The focus is on loudness, dynamics, spectral balance, and stereo behavior, based strictly on the provided metrics.

1. Phosphorescence_mix.wav

Loudness & Dynamics

• Original: -16.75 LUFS with high dynamic range (14.07 dB).
• TENSORMIX: Moderate loudness increase (-14.52 LUFS) while retaining relatively higher dynamics (12.54 dB) compared to other mastering services.
• LANDR / eMastered: Slightly louder (-14.26 / -14.12 LUFS) with more dynamic reduction (~11.4 dB).
• SoundCloud: Effectively unchanged loudness (-17.22 LUFS), closest to the original dynamics.

Spectral Balance

• TENSORMIX: Noticeable low-mid and mid lift with controlled highs; spectral center shifts upward moderately (2874 Hz).
• LANDR: More conservative tonal shaping, closer to original balance.
• eMastered: Stronger tilt and higher spectral center (3011 Hz), indicating a brighter, more forward tonality.
• SoundCloud: Minimal tonal change.

Stereo / Correlation

• All services remain within safe correlation values.
• TENSORMIX shows slightly increased stereo engagement without instability.

Summary: TENSORMIX strikes a middle ground between loudness and dynamic preservation. eMastered pushes brightness more aggressively. LANDR is conservative. SoundCloud acts more as normalization than mastering.

2. I_Want_It.wav

Loudness & Dynamics

• Original: Very quiet (-19.73 LUFS) with wide dynamics (16.5 dB).
• TENSORMIX: Raises loudness to -15.02 LUFS while keeping dynamics higher (13.56 dB) than competitors.
• LANDR / eMastered: Louder targets (-14.05 / -13.7 LUFS) with greater dynamic compression (≈12 dB).
• SoundCloud: Minimal loudness increase, dynamics largely intact.

Spectral Balance

• TENSORMIX: Maintains low-end stability and avoids excessive high-frequency attenuation; tilt remains negative but controlled.
• LANDR: Slightly scooped mids and darker highs.
• eMastered: Stronger mid/high attenuation, resulting in a darker presentation.
• SoundCloud: Closest to original tonal balance.

Stereo / Correlation

• TENSORMIX shows reduced correlation (0.57), implying wider stereo processing compared to others.
• Others remain near original stereo behavior.

Summary: TENSORMIX prioritizes dynamic retention with a moderate loudness target and wider stereo image. LANDR and eMastered push louder at the expense of dynamics. SoundCloud again behaves conservatively.

3. Drown_mix.wav

Loudness & Dynamics

• Original: Very dynamic (-21.33 LUFS, 18.8 dB).
• TENSORMIX: Raises loudness to -16.31 LUFS while preserving more dynamics (14.39 dB) than competitors.
• LANDR: Aggressive loudness (-12.81 LUFS) with significant dynamic reduction (9.65 dB).
• eMastered: Moderate loudness (-14.5 LUFS) with strong dynamic reduction.
• SoundCloud: Light loudness increase with dynamics largely preserved.

Spectral Balance

• TENSORMIX: Balanced tonal lift with restrained high-frequency emphasis; spectral center increases moderately.
• LANDR: Heavier mid and high attenuation despite high loudness, suggesting stronger limiting/compression.
• eMastered: Very high tilt (4.2) and highest spectral center (3190 Hz), indicating aggressive brightness.
• SoundCloud: Minimal tonal deviation.

Stereo / Correlation

• TENSORMIX shows the highest stereo correlation shift (-13.14 dB), indicating the widest stereo treatment among the services.
• Others remain closer to the original stereo field.

Summary: LANDR is the most aggressive in loudness, with clear dynamic sacrifice. eMastered emphasizes brightness. TENSORMIX balances loudness increase with dynamic and tonal restraint while expanding stereo width.

Cross-Track Observations

TENSORMIX

• Consistently targets moderate loudness.
• Preserves more dynamic range than other mastering services.
• Applies noticeable but controlled stereo widening.
• Avoids extreme spectral tilt.

LANDR

• Prioritizes loudness.
• Most aggressive dynamic reduction, especially evident on Drown.
• Tonal shaping is generally conservative but compressed.

eMastered

• Tends toward brighter, more forward tonal balance.
• Dynamic reduction is significant but less extreme than LANDR in some cases.
• Strong spectral tilt on certain tracks.

SoundCloud

• Functions primarily as normalization rather than creative mastering.
• Minimal tonal, dynamic, or stereo alteration.
• Closest to original mixes across all tracks.

Overall Conclusion

From a purely technical standpoint, TENSORMIX emphasizes balance: moderate loudness increases, comparatively higher dynamic retention, controlled tonal shaping, and wider stereo presentation. LANDR favors competitive loudness with notable dynamic trade-offs. eMastered leans toward brightness and presence. SoundCloud remains the least transformative and should not be considered a mastering solution in the same sense as the others.

(End ChatGPT analysis)

Final thoughts

I said in the beginning that I didn’t want to try to say that Tensormix is better than others, and I hope that this data can support that - because if there’s one thing that becomes obvious from looking at the numbers, is that all of the different services do things differently. LANDR has been around for years, and I remember using it when it first came out - it’s always been a service that catered primarily to electronic producers and as such, it’s natural that they prioritise loudness. Even though the loudness wars seem to have calmed down a bit in recent years, probably driven by streaming platforms’ loudness normalisation, loudness is still synonymous with EDM production.

But I am certainly happy with how Tensormix turned out - by no means is it meant to replace human engineers, or claim to be the best thing since sliced bread.

It was born out of a nerdy “What if” statement, and I hope to continue expanding on this nerdy foundation for a while.

EDIT (30/01/2026): Product renamed to Tensormix (previously Loudlink)