REW - Generated Correction Filter Impulse File - Length vs Quality?

ola1

Member
Thread Starter
Joined
Sep 27, 2017
Messages
21
Background

1. In my recent efforts, I exported the correction filter (further to its automatically creating 20 EQ's using the Generic device), generated by REW to a .wav file. 32 bit float, 44.1Khz.
2. Imported the correction filter file into a convolution plugin (VST), in my DAW - Reaper. on Windows 10.
3. And I can pass audio through this plugin to my speakers.

Observations -

I think the outcome of audio playback, in recent versions (I'm on REW 5.17 Beta 12)through the auto generated filter from REW represents the most balanced result, I have ever listened to. and I must thank the Author for this. I have used other methods like DRC Designer(which is a front-end to DRC), but this approach to generating a correction, while relatively simple, ese seem to need quite a bit of expertise to avoid artefacts. Truth be told, good is only realised when you meet even better. My recent use of REW to generate filter was purely by accident - just to see what it was capable of.

What led me to use DRC Designer prior to now, after REW for measuring, were limitations in the quality of correction files generated by REW about 2 or 3 years ago, where I could clearly see from importing the file back into REW that it only corrected frequencies no higher than about 8Khz, and at the time to seemed to be a relatively broad correction. My initial limited results with REW's auto correction/generation of correction filter impulses, could also have been user error, as one's initial efforts at the concept of DRC (the approach) in general, may take a while to fully appreciate.

Preamble over.

I observe that the file length of the exported filter file which I import into my convolution plugin is 2.972 seconds.

Questions

1. Is the length of this exported filter file fixed at 2.972 seconds, or does it vary based on the number of filters and the parameters of these filters generated by REW?

2. I observe that there is some blank samples(silence I think), before the impulse itself starts, in the exported file... Is there any reason for this? Is there any benefit or disbenefit of my manually deleting/trimming this blank silence before the impulse, to reduce this latency of one or two samples of silence which this, in theory, introduces?.

For most playback it is obviously not an issue in practice, not something that anyone would notice, I was only thinking aloud.

3. Within the generated file, there are broadly speaking 4 consecutive sections, which I attempt to describe below.

a. The silence at the start, for a few samples. Exactly 2 samples of silence if I am correct (or my audio editor Goldwave - is correct).
b. The impulse of the correction filter, which ends by about 10 milliseconds. By about 5 milliseconds, though from visual inspection, there's not much information left...
c. From 10 to about 150 milliseconds, there even less information in the file, but above -90dB.
d. Beyond 150 milliseconds till the end of the file @ 2.972 seconds, it's pretty much silence, definitely below -90 dB..

What is the minimum length (in time, e.g milliseconds), of the filter impulse file that provides the maximum audio quality, In the convolution result? And does this minimum length of the file change depending ?

Another way of asking this is how much of the relatively low volume or silence at the end of the file is required to be in the file, in case I wanted to experiment and truncate this?

I ask because my experiments with truncating this silence beyond about 100 milliseconds, when I do an AB test with the full length file reveals a more up front presentation for the trimmed version. - Crisper - more in your face. While the full length file almost feels like the audio is very slightly recessed in the speaker - slightly woolier, like its coming from one or two feet inside the speaker. Most people would not notice any difference, but I did, which is why I'm here asking.

I still cannot figure out if the shorter trimmed version of the correction filter is the better one, but things like transients, guitar plucks are more - in your face, voices like Adele's 19 album seem pretty much like the microphone is almost in her throat - especially lead voices in the "silence trimmed" correction filter file are that it more lifelike, and every nuance such as words, are more discernible - both for sung vocals or recorded speech such as documentaries. I almost wonder - by trimming, have I added distortion to the end result, as a by product and this makes it sound better, or am i actually hearing a better result from this trimmed version.

Why does the perceived audio quality change with the length of the filter impulse file? I am definitely hearing a difference - definitely - trust me. I've done blind testing, by randomly switching between two instances of my convolution plugins set to different file length versions of the filter impulse file, which I am able to switch between, using a hot key, and I can reliably tell which impulse is which, by listening. And you can test this for yourself.

To the Author of REW, John, you have been most generous, and how ever many times I say thanks, it will never be ample, you generosity has made a huge difference in how I experience audio.

My reciprocal must be to share something really important that I discovered through critical listening, which I have not heard or read discussed anywhere. My own way of giving back.

The convolution playback engine has a direct impact on the audio quality - you can test this yourself, you'll hear it eventually, or maybe immediately if you have very high resolving speakers, or your correction is excellent. I was shocked at this - thinking that well convolution must sound the same, being of the opinion that this was a mature science, but there are differences between products.

1. I tried one of the Liquidsonics products, the free one - it's great.

2. Then its a toss up between the convolution in Reaper's Reaverb which is also available as a free unbundled plugin, within their Reaplugs free bundle of plugins) for use in any other DAW, and IR-1/IR-L from Waves (not free - I own both - but also not too expensive, if you buy when they are on sale - the waves products have more bells and whistles) subjectively the quality of audio between these is about the same, but the sound is not identical.

3. The product that consistently gives me the most accurate result, IMHO, after extensive testing over at least a year is the Impulser2 convolution plugin which is a component of the Freeverb3 bundle (freeware). This uses, double precision, which I think is probably the key to its accuracy, and it is delivered with both versions - one fast (single) and the other a slow (double precision) for Windows (Mac version IIRC has only the single precision - but I do not use Mac so please check this - ). For speaker correction purposes, using only 1 instance, the additional CPU overhead on any Intel or AMD based CPU released in the most recent 8 years is negligible. That's the best I can do, highlight one of the most important tweaks to improving the result of DRC - the convolution payback engine, and this one is free, yet higher quality than payware.... Within a few hours of critical listening - actually more like a few minutes probably you hear the difference.

If I may add, rather than trimming the length of the file as discussed above, most of the convolution tools provide a method of defining what portion of the filter will be used - start time, end time, so you can achieve similar results/tests using the full length export from REW directly, without having to actually create a trimmed copy of the filter file.

I really look forward to understanding the aforementioned phenomena much better, and hearing from others about this.
 

ola1

Member
Thread Starter
Joined
Sep 27, 2017
Messages
21
I did a further test - this time pretty objective - routing an audio track 1 through track 2 and 3.

Track 2 has the convolution plugin using the full length correction filter - impulse file

Track 3 has the convolution plugin using the trimmed correction filter - impulse file

Then track 2 and 3 are routed to track 4 where I have an analyser, image should be below -

Blue line = convolved with full length file

Pink Line = convolved with "silence trimmed" file

Clearly using the file with silence trimmed, changes the way the low end sounds, a tad..., and its audible.. So I was not imagining this difference.

The crest factor is also different, the boosted low end on the untrimmed convolution file, seems to influence the measured dynamic range, but in reality the trimmed file convolution sounds more dynamic, cos with the reduced low end, the mid range is more audible.

Whats instructive is that there is very little difference in the higher frequencies, at least not visible in this analysis.


Filter length difference.png

The image (hopefully attached)
 

John Mulcahy

REW Author
Joined
Apr 3, 2017
Messages
7,212
Quite the first post!

The EQ filters are of the IIR variety, so infinite impulse response in theory. In practice the responses decay fairly quickly, the lower the filter's frequency the slower the decay. The export is chosen to have at least 2 seconds of data so there is sufficient time for the impulse response to decay below the resolution of most audio data. The length will only depend on the sample rate. Truncating the response alters the filtering, but truncating it after it decays below the resolution of the audio data would have little measurable effect.

The 2 zero samples at the start were to help a little with windowing the response if it was re-imported, but that has been better managed since 5.20 beta 6:
  • Exported filter impulse responses now start at sample zero rather than sample 2
  • Imported impulse responses are padded with additional zero samples before and after the response if the peak is near the start of the response
 

ola1

Member
Thread Starter
Joined
Sep 27, 2017
Messages
21
John M,

Truly appreciate your response. Thanks for clarifying my understanding.

I did some more listening, after a short break, and at this time I am convinced that the "brightness"/"clarity" I was hearing earlier, when audio was processed with my manually "trimmed" filters, is artificial/an artefact/an effect of an incomplete filter.

Experimentation over. I am absolutely happy with using the generated filter as is.

Using Impulser2, and your(REW) correction filters - default export - no changes by me, this is the most authentic reproduction of audio I have ever heard, on any sound system. Especially when listening to music that has been recorded and most likely processed digitally, at very high fidelity, yet sometimes the words/lyrics become difficult to understand, just comes out now - pin sharp, like having a magnifying glass. Depth, Height, Clarity, at all volume levels. Another level of human experience - truly another level. And this clarity/3D illusion is maintained when listening to only one speaker. Awesome...

Happy to enjoy the limitations - sounds stunning like the speakers do not exist and music is coming from somewhere in the ether in front of the speakers, from the the ideal listening area, mine is about 2 meters x 2 meters = 4 meters square. - but that's perfect. More than happy to live with this. Yet it sounds spectacular almost anywhere else outside the listening area, and especially when one recognises that the higher frequencies are most directional and will diminish off axis, like pretty much every other speaker, the off axis response is all things considered still amazing.

One caveat, which I only discovered a few minutes ago, with the Impulser2 convolution plugin(VST). It has a bug, in the display of its limiter button, so it is difficult to know when it's on or off, always resorts to displaying that the limiter is on - so my workaround has been to set the Threshold and Ceiling to their minimum values of -18.5, so that if the limiter is turned on, I will hear this immediately with most pop music, as distortion in the audio stream, and turn it off. Hopefully I'll reach out to the developer and kindly request a fix for this bug.

For anyone who may wish to reference some music that will challenge the system from a resolution standpoint, this will surprise many - Beyonce's eponymous album from 2014, has some spectacular sonics, it was clearly recorded with great converters, and superb processing, which you hear clearly once the correction filters are "turned on". - dare I say it again 3D, might not be everyone's taste lyrically or genre wise, not what I also listen to a lot, but it demonstrates what a sound system can become when it becomes highly resolving.

Glad that I can now work in one tool for measuring and generating the correction filter(impulse). REW.
 
Top Bottom