[Rayon]

I've seen some discussion in both Susvara and Abyss 1266 thread, but I think it's worth it's own thread.

Long story short: Canadian company Accurate Sound makes convolution filters that one can use in Roon and HQPlayer to flatten the frequency response curve of headphones. It also fixes problems with phase.

Examples:

Susvara

Abyss 1266 Phi TC

I bought this filter for Susvara and it has been very positive experience. I have moments when I want to use Susvara with filters and sometimes I listen to it just as is. I feel that this has given me another pair of Susvara with completely different sound signature and I feel that sound quality is very high. I've also received A+ service from Mitch when I've had questions regarding to filters. These filters are not cheap, but neither are tubes and cables that people buy to use in fine tuning and the effect that this filter makes in sound is huge. For me there's value in owning flat tuned headphones for reference purposes as well.

Link to their web shop: https://accuratesound.ca/store/#filter

 

[pk4425]

What are the advantages of these convolution filters over the ones that Jaakko Pasanen offers for free for hundreds of headphones? Sincere question.

 

[Rayon]

What are the advantages of these convolution filters over the ones that Jaakko Pasanen offers for free for hundreds of headphones? Sincere question.

At least phase correction. Then these are flat while Jaakko's are Harman target.

 

[pk4425]

At least phase correction. Then these are flat while Jaakko's are Harman target.

Ah, interesting -- and a key difference. I'm a bit slow on the uptake! Thanks.

 

[GoldenSound]

At least phase correction. Then these are flat while Jaakko's are Harman target.

EQ will also correct phase, FR and phase are inherently linked

 

[Rayon]

EQ will also correct phase, FR and phase are inherently linked

EQ will affect phase, but if we use EQ to make headphone's frequency reaponse flat, does that also always guarantee flat phase reaponse? To me it sounded like Mitch's filters would do something extra to remove Susvara's phase anomalies.

 

[GoldenSound]

EQ will affect phase, but if we use EQ to make headphone's frequency reaponse flat, does that also always guarantee flat phase reaponse? To me it sounded like Mitch's filters would do something extra to remove Susvara's phase anomalies.

Yes
So long as the actual adjustments being made by convolution and FR are the same then you'll end up with the same FR and phase response.

Convolution isn't actually doing anything fundamentally different to EQ. It's just that in situations where you're needing to make rather complex adjustments a convolution can be much easier than setting up an EQ profile with 50 bands etc. But you can indeed do it either way.

Additionally it's worth noting that there are also big drawbacks to trying to EQ something flat. The biggest one being that to have the intended effect it requires that your HRTF is identical to the measurement rig. As well as having identical placement. Which would be statistically near impossible.

As a result, the convolution will be mostly accurate upto 5-7khz or so. But above that and certainly for anything past 10khz it's just not going to be accurate at all and could very easily make things far worse.
Because of this some people create convolutions that don't alter stuff above 5khz or 10khz other than maybe an overall treble shelf instead of smaller peak/dip adjustments. From the images above it does not look like that's being done here so for many people these convolutions will end up sounding very bad.
Though could maybe ask the creator for ones that only go up to 10khz for a better result

 

[mitchba]

@GoldenOne , I agree with everything you are saying, however, I think you may have overlooked the high frequency response limit of the chart that @Rayon posted. The chart goes up to 3 kHz...

Accurate Sound's convolution filters do not eq HRTF. There is no eq beyond 6 kHz, other than like you say, perhaps a high frequency shelf that adjusts the overall treble level to one's ears.

A FIR filter with 65536 taps @ 48 kHz has a frequency resolution of 48000/65536 = 0.732 Hz. The frequency range spans 0 Hz to 24 kHz (fs/2). Thinking of a FIR filter as a graphic equalizer: 24000/0.732 = 32768 sliders for our equalizer. That is 1000 times the frequency resolution of a 31 band 1/3 octave equalizer.

Kind regards,
Mitch

 

[manishex]

@GoldenOne , I agree with everything you are saying, however, I think you may have overlooked the high frequency response limit of the chart that @Rayon posted. The chart goes up to 3 kHz...

Accurate Sound's convolution filters do not eq HRTF. There is no eq beyond 6 kHz, other than like you say, perhaps a high frequency shelf that adjusts the overall treble level to one's ears.

A FIR filter with 65536 taps @ 48 kHz has a frequency resolution of 48000/65536 = 0.732 Hz. The frequency range spans 0 Hz to 24 kHz (fs/2). Thinking of a FIR filter as a graphic equalizer: 24000/0.732 = 32768 sliders for our equalizer. That is 1000 times the frequency resolution of a 31 band 1/3 octave equalizer.

Kind regards,
Mitch

Any recommendations on a good basis for a high frequency shelf? One from 6khz or 10khz?

Would we benefit from having more Taps on the filter, or would the drawbacks outweigh this?

Since all headphones vary over 6khz, which is your favourite, most natural headphone whilst using the filters?

 

[mitchba]

Any recommendations on a good basis for a high frequency shelf? One from 6khz or 10khz?

Would we benefit from having more Taps on the filter, or would the drawbacks outweigh this?

Since all headphones vary over 6khz, which is your favourite, most natural headphone whilst using the filters?

Yes, I would start with a 6 kHz HF shelf and adjust a few dB up or down to taste.

A 65,536 tap FIR filter has more than enough frequency resolution. The low frequency limit of a FIR filter is typically 3 x the frequency resolution, which in the previous example is 0.732 Hz. So the filter is processing down to ~2 Hz. Longer tap filters could be used, but there is no audible benefit with the drawback of increasing CPU and memory utilization.

LCD-4

 

[Minimyna]

Sounds intriguing but $150 for a software filter seems a bit excessive. For $50 i might have tried it.

 

[GoldenSound]

Accurate Sound's convolution filters do not eq HRTF

HRTF isn't related solely to high frequencies though. I'm aware there is a HF limit, which is good as of course the response above 7khz or so varies so much from person to person that trying to EQ there based on either someone else's in-situ results or indeed the results from a HATS is a bad idea, but that doesn't mean the in-situ response will be the same for all people beneath that. This is one of the challenges with standardizing headphone measurements as there isn't even a way to really get a 'one size fits all' representation of a diffuse field let alone an in-situ device.

This is the HD800S measured on three HATS for example

As a result, your own HRTF if using in-ear mics (plus some possible alteration from the in-ear mics themselves, that's an entire rabbithole in and of itself with many challenges and blocked canal depth, 2nd bend influence etc, getting in-ear mics that will give you a result really close to what's actually at your ear drum is pretty hard and requires you to get very specialized custom made units from deep reaching molds) will be altering the result. This can be verified by just comparing the measured in-situ result to a decent HATS and observing the differences.

A FIR filter with 65536 taps @ 48 kHz has a frequency resolution of 48000/65536 = 0.732 Hz. The frequency range spans 0 Hz to 24 kHz (fs/2). Thinking of a FIR filter as a graphic equalizer: 24000/0.732 = 32768 sliders for our equalizer. That is 1000 times the frequency resolution of a 31 band 1/3 octave equalizer.

This is a comparison to a basic fixed graphic equalizer though, not a proper parametric equalizer with variable Q and Freq. In situations such as room correction where the very complex adjustments would require an enormous number of high Q filters to replicate, it means convolution is sort of the only way to go. But here, where the adjustments being made are much more simple (which is fine, they shouldn't be doing any high Q adjustments), that's not the case.

The image below for instance shows the response of a log chirp with your Susvara convolution filter applied and the other curve being a 13 band parametric EQ applied.

Generic
Number    Enabled    Control    Type    Frequency(Hz)    Gain(dB)    Q    Bandwidth(Hz)    TargetT60(ms)    FilterT60(ms)  
1    True    Auto    PK    50.80    -1.50    2.001    25.39  
2    True    Auto    PK    92.80    -1.80    2.000    46.40  
3    True    Auto    PK    170.5    -1.90    2.002    85.16  
4    True    Auto    PK    196.5    -0.50    49.854    3.94  
5    True    Auto    PK    365.0    -1.90    3.209    113.7  
6    True    Auto    PK    606.0    -1.50    5.413    112.0  
7    True    Auto    PK    1007    -1.80    1.021    986.3  
8    True    Auto    PK    1887    4.30    1.960    962.8  
9    True    Auto    PK    2245    1.60    2.380    943.3  
10    True    Auto    PK    3883    -5.80    1.504    2582  
11    True    Auto    PK    5251    -3.30    4.056    1295  
12    True    Auto    PK    11814    -1.00    1.209    9772  
13    True    Auto    PK    18126    -1.40    1.044    17362

The other reason of course to use convolution would be if you wanted to adjust phase independent of frequency response, but with headphones being minimum phase that isn't something we want to do.

 

[EMINENT]

Hielo.

I and others have tried a similar rough trace and it did not sound the same. I would be happy to test another trace again.

 

[GoldenSound]

Hielo.

I and others have tried a similar rough trace and it did not sound the same. I would be happy to test another trace again.

What was the eq setting you used/how was it calculated?

 

[EMINENT]

What was the eq setting you used/how was it calculated?

Huilo.

From here: https://forum.headphones.com/t/audeze-lcd-5-official-thread/14588/784

Granted, it does not look the same as it is just a copy or rough trace.