Skip to main content
Topic: ABXing low-pass (Read 4723 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

ABXing low-pass

To support my claim made in

The cymbal included in the EBU test data was played too softly and there was no hi-hat, so I used the harpsichord sample, which I cut and low-passed using SOX (sox harp-cut.wav harp-16k.wav lowpass 16000)

I was able to ABX a 16kHz lowpass on my modded HD238 headphones and my Acer laptop (HDAUDIO\FUNC_01&VEN_10EC&DEV_0269&SUBSYS_10250364&REV_1001, ).

I wasn't trying to match A to X/Y but I was merely picking the non-low-passed version from A versus B, then X versus Y. The ABX process itself was fatiguing as hell, I was already exhausted around number 8.

ABXing low-pass

Reply #1
SoX's "lowpass" effect is a Butterworth filter, which has too gradual a roll-off to be useful for something like this. It attenuates 16kHz by 3 dB, and is basically 12dB per octave above the cutoff point, so 20kHz is attenuated by only 7dB (3+12*log2(20/16)). Below the cutoff it's more gradual, so e.g. 15dB would be attenuated by almost 2.5 dB and even 12kHz is attenuated by ~1.2dB. So it's possible that your ABX was due to the effect it was having on the sub-16kHz spectrum. Use 'sox in.wav out.wav sinc -16k' instead.

ABXing low-pass

Reply #2
Yeah, it’s very important to realise that all LPFs are not equivalent, so you can’t just say you hear a difference with one filter and assume/imply that applies to all of them.

In order to be able to have any confidence whatsoever in which cutoffs you can or cannot hear, you’re going to need to use an LPF with a steep slope and hence minimal attenuation within the passband. But often, the steeper the slope, the more chance of ripple or greater disturbance to phase within the passband, things that might create audible differences of their own, which would confound the test.

However, as jensend indicated, there are certainly better filters to try in the first instance; a Butterworth isn’t informative.

ABXing low-pass

Reply #4
Based on the equally loud (if not louder) content in the adjacent octave just below 16kHz [I visually see using FB2K's musical spectrum component, right clicking it, and in "Options" expanding the high frequency cut-off note (frequency): ] in your full-bandwidth harpsichord sample, and assuming you aren't keying off of any of the possible artifacts db1989 mentioned above, it would appear to me that you are immune to masking as I understood it, especially considering how loud frequencies at >16kHz have to be in the first place (to be audible, due to equal loudness curves), ignoring the prescence of a "masker" which makes the threshold of audibility for those top frequencies even higher.

P.S. Can anyone find me a version of these following images, or at least similar, but with the masker at about 8-16 kHz? [Everything I could google up stopped at 8k]

Wikipedia source article:

[This second image better shows how notably louder >16kHz sounds have to be in the first place. It's not exactly what I'm seeing in the harpsichord high frequency levels when viewing the spectrum analysis.]

ABXing low-pass

Reply #5
The lowpass makes a tiny difference in the attacks, especially for the second keystroke. That's what I was "looking at" to tell the samples apart.

I'm not immune to masking, I'm entirely happy with a 192kbit-256kbit AAC most of the time. I think that whoever came up with the above data simply used "normal" people (as in people not trained for critical listening, no audiophiles etc.).

ABXing low-pass

Reply #6
Is it possible that the lowpass filter changes how the encoder treats other (audible) bands?

Edit: Never mind, I didn't notice that the samples were lossless.

ABXing low-pass

Reply #7
I was ABXing PCM wave files.

SimplePortal 1.0.0 RC1 © 2008-2020