Mar 23, 2023 – initial publish
Mar 23, 2023 – added Sibilance, Pro-DS, and 902
This is not a scientific test, and these results should be considered experimental at best. YMMV, IANAL, results may vary…
There are a lot of different de-esser plugins out there. As an audiobook narrator, it can be overwhelming to try to figure out which is the best tool to use. So, I decided to do some comparisons and work out some of the details. The results were useful to me, so I thought I would share.
My go-to plugin is iZotope RX De-Esser, which comes as part of the RX Standard bundle (which typically retails for $150 with crossgrade pricing during sale season). My goal in these tests was to find out if there was another plugin out there that could do a comparable job to iZotope De-Ess, but do it faster (i.e. with less CPU consumption).
[Ed. note: In the course of these tests, it became clear that the +/- efficiency was not a significant factor in most cases – discarding the top and bottom results, a spread of only 14% is observed. This is effectively a rounding error given how unconstrained my test environment is. Therefore, for the majority of use cases, plugin speed should not be given undue importance when choosing a de-esser for your workflow. These will all perform relatively equally.]
To make sure that the results were as consistent as reasonably possible, I opened up a random recent chapter recording that was about 10 minutes long, which had moderate sibilance in a couple of places. Rendering this audio would give us a reasonably consistent gauge of average render time. For each test case, I rendered only the delta-solo output of the de-esser plugin, so I could see how clean it was working, and analyze the results. This tells us the approximate efficiency of the plugin.
Once I created a wav file containing only the removed sibilance, I then pulled this into iZotope RX Audio Editor, and took an FFT average. These graphs show us what the de-esser is actually removing (this is important later).
For all plugins, I attempted to keep settings as similar as possible. Not all plugins expose parameters like attack and release, but parameters like threshold and frequency are usually accessible.
[Ed. note: After releasing these results, some wiser heads explained how these test methods are not as useful as I’d hoped, because not all de-essers function the same way. Some apply broadband compression to the entire signal based on the target frequency; others operate as multiband, or dynamic EQ and only affect a smaller and more precise, respectively, portion of the overall signal. Some algorithms implement a “harshness” approach to sibilance detection, which again cannot be directly compared on an FFT analysis to what other implementations do. That said, the following analyses are left here for anecdotal and inspirational purposes only, and should not be used as a reference for anything other than painting mountain landscapes.]
PLUGIN CANDIDATES AND RESULTS
For this round, I compared nine popular plugins based on common recommendations. For simplicity, I’ll summarize the test results in a table here, with individual spectral and FFT results below for those who are curious.
iZotope RX 7 De-Ess
TDR Nova (Gentlemen’s Edition)
DMG Audio TrackDS
Lindell 902 De-esser
Free / GE $10 (sale)
Free / Pro $90
Click for 1:1 view
I will post individual FFT analytics and specific plugin settings for each of these tracks, but first a few bird’s-eye observations from just the spectral view in Reaper of all the tracks at once:
- iZotope is removing the least amount of information overall. T-De-Esser is a close second.
- TDR Nova is hitting the same sibilances, but is also working on a lot of other stuff that’s not sibilance. You might be able to dial this in more tightly. However, it’s 20% slower than iZotope already, so it’s a hard pass from me from the start.
- Techivation, AO, and TrackDS are doing a solid job, and are picking up some very small sibilances that iZotope is not. They both are hitting more -t and -d sounds (not just -s) that are a little chirpy. I would typically not worry about these as much, but they’re obviously sharing the same bandwidth.
Just from these comparisons, here’s the high-level conclusion:
- If you absolutely need speed, bite the bullet and shell out for Waves De-Esser (best on sale), or iZotope.
- If you don’t care about render performance and just want the most transparent de-sibilance, buy TrackDS.
- If you’re concerned about cost and can take a small performance hit, buy the Pro version of Techivation or use the free version if you don’t care about the extra bells and whistles.
- If you just need the cheapest option that does a reasonably good job, subscribe to the AudioObsession Patreon for a month so you can get LOADES for $5. Techivation’s T-De-Esser is free, but is a little less transparent than LOADES.
For those who want the more detailed analysis, next I’ll show the FFT analysis for each of the delta-solo tracks, with the corresponding plugin settings used in each case.
This sample audio was recorded in my treated home studio with a Blue Spark SL Blackout into a Motu M4. In the following results, you’ll see a dip at 9.8k that is a product of the frequency response of this microphone.
iZotope RX 7 De-Ess (Spectral Mode)
Classic mode, shown above, is described as a broadband attenuator and will certainly affect lower frequencies. I re-ran the test with it in Spectral mode, adjusted the Threshhold and Cutoff frequency from -25/2500hz to -16.4/6100hz respectively (this most closely matched the previous results), and left the rest of the settings default (Slow mode, 50% aka neutral shaping, no spectral tilt).
iZotope RX 7 De-Ess (Spectral + Tilt)
In a mad frenzy to mitigate that low-end damage, I increased the
Spectral tilt function 50% towards the White noise profile, and re-processed. From the De-ess documentation:
Spectral Tilt gives you the flexibility to determine what the ideal shape of your high frequency signal could be. Moving the shape toward brown noise moves it toward a darker sound. Moving the shape toward white noise provides a brighter result. Together with flattening and threshold, you can determine how much shaping takes place.
We can observe some reduction in low-frequency removal, but still quite a bit compared to LOADES and TrackDS.
TDR Nova (Gentleman’s Edition optional)
As noted before, since Nova is using a broadband and set-frequency EQ, it’s bleeding into a lot of other things in an undesirable way. You may be able to fine-tune this; the GE edition provides a “surgical” mode which may be helpful. However, this plugin still ran 20% slower than the benchmark so I determined it was not worth pursuing – especially since there are other, easier, more transparent, free plugins available.
The FFT curve here is remarkably similar to iZotope’s, as previously noted. Since the standard version of this plugin is free, this makes it a strong contender (and is only about 7% slower).
At the time of writing I noticed that there is a significant update to the plugin available (version 1.1.15) – this may be worth re-testing. Apparently the default download link is pinned to the older version.
no version number
AO’s plugins are available for anyone subscribing to the Patreon; the minimum tier is $5/mo (currently offering 50 plugins, so a ridiculous value overall). It’s tilted towards the 6-10khz band much more than T-De-Esser, but is doing almost nothing below 3khz, which is interesting.
DMG Audio TrackDS
TrackDS makes a case for being the most transparent of all the plugins, alongside LOADES. It also has some useful fine-tuning controls, which LOADES lacks. Interestingly, turning off the highpass and dropping the frequency threshold from 5.3k to 3k increased render speed from 5.8x to 6.1x realtime.
I think this implementation is largely overkill for narration applications, given the variety of exposed parameters like sidechaining, MIDI learning, stereo linking, etc. However, the ability to control lookahead and oversampling (up to 4x) is nice. I don’t love the relatively huge low-end reduction tho.
This test was never intended to be an empirical head-to-head, nor a definitive reference. However, it may provide a little bit of context and a data point for those who are trying to wade through the sibilance tamers available out there.