Hacker News Reader

I was listening to some of the LAME quality test samples[0] and I found that the hi-hat pre-echo test sounds worse at 160kbps than at 128kbps (LAME 3.100, -q 0). I thought it might be because of the different low-pass filter frequency, but even with manually specified matching lowpass (16805Hz) for the 160kbps version the pre-echo sounds worse.

The 128kbps version doesn't sound much like the original, but it's still a pleasant sound with hardly any pre-echo. The 160kbps version, even with the extra lowpass, has obvious and annoying pre-echo on the first hit.

[0] https://lame.sourceforge.io/quality.php

I can confirm this. I tested with 3.100.1 and the following encodes of hihat.wav:

    -q0 -b128 (b128)
    -q0 -b160 (b160)
    --preset cbr 128 (pc128)
    --preset cbr 160 (pc160)
    --preset 128 (pa128)
    --preset 160 (pa160)

Using the Foobar2000 ABX comparator I was able to 5/5 ABX:

    wav vs. b128
    wav vs. b160
    b128 vs. b160 (this one was 7/8 ABX)
    wav vs. pa128
    wav vs. pa160
    b128 vs. pa128 (more difficult)
    pa128 vs. pa160 (more difficult)

I did not successfully ABX b160 vs. pc160.

I didn't do ABC/HR so I can't say for sure what my ranking would be, but my guess is:

wav >> b128 > (b160 or pa160) > pa128

I did an investigation into MP3 vs AAC. My tests were at 320kbps.

https://andrewrondeau.com/blog/archive/2016-07

Basically, MP3 is crap at 320kbps, and AAC is somewhat better.

The problem is that both formats roll off high frequencies. It's fine if you're listening in a car, through cheap speakers, on a noisy plane, aren't sensitive to higher frequencies, ect, ect. But, if you're in a quiet room with good speakers / headphones, and you have good ears, the high frequency roll off is noticeable.

This appears to be visual comparisons of graphs of various automated measurements. Lossy codecs are designed to exploit weaknesses of human hearing, so the only way to test if they're working as intended is with listening tests. This is most conveniently done with ABX testing software. If that high frequency roll off is really noticeable you should be able to produce a statistically significant result.

Here's an online ABX test of 320kbps MP3 using a modern encoder:

http://abx.digitalfeed.net/lame.320.html

Very few people are capable of hearing the difference, and even then only in difficult-to-encode samples. I don't think it's reasonable to call something of this quality "crap".

> Very few people are capable of hearing the difference, ... I don't think it's reasonable to call something of this quality "crap".

But some people can tell the difference. Therefore, the roll off is crap.

Granted, if you only give yourself ~3.5 bits / sample, MP3 is quite impressive for what it does.

The audible difference is mostly pre-echo and warbling noises, not low-pass filtering.

"Investigations" without blind tests are useless. Try a blind test (the website of the sibling post is very interesting) and you'll be surprised at the bitrate at which you'll fail to notice differences.

Did you consider re-testing 16bit vs 24bit?

HE-AACv2 is better at low bitrates (spoken podcasts) then mp3. ffmpeg -i input.wav -c:a libfdk_aac -vbr 3 -b:a 32k -profile:a aac_he_v2 output.m4a

There is also xHE-AAC https://gitlab.com/ecodis/exhale https://www.mainconcept.com/hubfs/PDFs/User%20Guides/MainCon...

Most encoders apply a low pass filter to cut off frequencies above 16kHz. Some do it above 18kHz but most choose 16kHz as default setting. This is common in all lossy codecs - mp3, aac etc.

It depends on the bitrate.

AFAIR, LAME applies at 16 KHz cutoff at 128 kbps, and 17 Khz at 160 kbps. I don't have data about the others, but AFAIK a similar approach should apply.

I'm not able to hear above 16 Khz. I've tried only once an experiment with another person, and they were definitely able to discern the difference.

>Something I remember is that MP3 spends a disproportionate amount of storage in order to encode frequencies > 16 Khz. This may explain why in some tests, it performs worse in mid-high bitrates (160/192 Kbps), although I don't know the technical details.

I wonder if just straight out filtering them would result in better 160Kbps MP3s, it's not like most people can hear them...

LAME does by default.

MP3 above 16Khz is the sfb21 ( scale factor band 21) problem.. Can’t be solved without lots of bits being thrown at it..

Thank you, that's very interesting. Are more modern format less susceptible to this problem?