Me vs. LANDR
(originally published at the Null Device Blog)
LANDR, an online automatic mastering service, just announced a major upgrade to their service. Their claim is that it provides “more focus, color and dynamics” and “rivals professional engineers.”
That’s a bold claim. A claim that cannot go unchallenged.
Brian Hazard of Resonance Mastering already did a pretty in-depth comparison of their early version. I thought it might be interesting to try it with this new version and see what’s improved.
My tests aren’t scientific, were pretty roughly tossed-off in a short amount of time, and have a pretty small sample size, so I can’t call them definitive in any way. But since LANDR gives any new signup two free “low resolution” downloads of mastered tracks, that seemed to be enough for a quick check of their claims.
The tl;dr version: pro mastering engineers don’t have to start worrying yet. It’s a definite improvement, and better than bad mastering, but it’s not exactly good mastering either.
I tossed our track “Wardrobe” at LANDR to see what would happen. It’s a track with a fair amount of dynamic range, a fairly open mix, and it has a bit of nuance that I thought might better display what LANDR does and does not do – a banging EDM track would probably sound fine no matter what, since those tend to be low on subtle dynamic shifts.
I have presented it with two versions – one with my usual array of 2bus post-processing (a SSL-style compressor and a tape sim), and one with no bus processing whatsoever. I also bounced one that I quickly mastered myself. To make it a fair comparison I exported mine as a 192kbps mp3 like theirs. Everything sourced form 24bit/44.1khz wavs. I used their guidelines for 6db of headroom.
LANDR gives you choices for “intensity”, which as far as I can tell means “how hard it hits the final stage limiter/clipper.” I selected “medium” intensity for the song, since it seemed a good fit. I previewed both low and high intensity as well, and while low I think gave better results (high was awfully overcompressed) it was also barely distinguishable from the original unmastered version.
On the plus side, it handled my bus-processed and unprocessed versions fairly identically. You can hear a bit of the missing saturation and some of the glue that the buscomp brings in, but the output is pretty consistent. I was interested primarily in seeing if mixing with a buscomp (as many engineers do) would mess up its algorithms. It does not appear to have trouble there.
In the end it turns out I pushed my own master slightly harder than LANDR did theirs. This is evident in some of the analysis I did.
Some things worked pretty much as expected. Both LANDR versions came in around -11/-12 LUFS (EBU R128+18), and both were hovering around a mean RMS of +4db (K-14). Unsurprisingly, the version without the bus processing had the highest crest factor and dynamic range of the three, while the bus processed versions (mine and LANDR’s) came in pretty closely – mine was about 1.5db less, but I also pushed the whole thing into the limiter about 2db harder while trying to match the loudness of LANDR’s preview.
The lower bitrate mp3 does it no favors. But aside from that, there are a few issues. There’s a definite harshness to the upper midrange. It did provide “focus” to the vocals, but almost too much – it feels like it added a big presence bump to the thing around 1500hz or so. Snares and cymbals get similarly “esh”-y. It’s less of a problem on the one without any bus processing, but it’s still noticeable. There feels like an overall bias to the midrange in general; I suspect the “genre detection” identified this as “pop” and gave it a radio-friendly curve with little bass boost or upper enhancement.
More worryingly, even with the >6db of headroom I gave the tracks, true-peak metering clearly showed some intersample overages (one as high as .7db), and a hard limit right up to 0db. While there wasn’t any audible distortion, that’s a pretty sketchy line for a professional service to walk. It might be an artifact of the low-res mp3 conversion process, but even that’s easily avoided.
Overall, it did a passable job, but it lacked some of the smoothness and balance I would normally expect from a pro master, and certainly didn’t do the same job of identifying “trouble spots” in the mix that I would normally manually compensate for (or send back to be re-mixed). Mastering engineers probably don’t have much to worry about yet. I can’t see this cutting into their business – the kind of person who would pay $4.99 to master a track probably wouldn’t have gone to even a $20ish/track mastering engineer to start with. The intensity setting seems to give some control, but it seems that neither high nor low are particularly useful. I suppose “high” is aimed at the “slamming EDM” crowd, but it doesn’t feel especially useful outside that genre.
Given more time and a willingness to spend some money on this (not quite to that point yet) I’d like to try this on a difficult track, one with mix problems or stereo width issues, or even possibly phase problems, and see if it’s capable of fixing/compensating for them. My guess is probably not.
I didn’t have enough of a sample size to try it, but given the original purpose of mastering is to prepare everything for a final media, I wonder how well it would do on a whole album? The trickiest bit of mastering, in my experience, is balancing multiple disparate tracks from disparate styles across and entire album. I can’t predict how well LANDR would handle that – but certainly I wouldn’t want it treating my ballads the same way it does my dance tracks and vice versa.
A double-blind test would really be the way to go with this, but I’m not quite to that point yet.
Interesting, at any rate.
Version 1 (bus processed, LANDR-ized)
Version 2 (no bus processing, LANDR-ized)
Version 3 (bus processing, my master)