Lossy encoders rarely devote many bits to high frequencies anyway.

I'm evaluating Opus for internet radio purposes (@72kbps and @160kbps) and wondering if feeding a 16kHz lowpassed signal to the encoder would give worse, same or better quality than encoding a full band 20kHz signal? The lowpass actually comes from my multiband compressor and i don't really like to keep content above 16kHz because it seems that >16kHz content may stear the multiband compressor away (some EDM have loud signals in that region).

Quote from: darkbyte on 2017-02-06 19:25:10

I'm evaluating Opus for internet radio purposes (@72kbps and @160kbps) and wondering if feeding a 16kHz lowpassed signal to the encoder would give worse, same or better quality than encoding a full band 20kHz signal? The lowpass actually comes from my multiband compressor and i don't really like to keep content above 16kHz because it seems that >16kHz content may stear the multiband compressor away (some EDM have loud signals in that region).

Low-passing is very unlikely to give better result, but it *probably* won't be much worse either. I've never actually experimented with that, so you should just try and see for yourself.

When I was feeding Opus at quite low bitrate with lowpassed signal, Opus treated high frequencies badly. If the lowpass frequency didn't match Opus encoding bands' transition, the encoder produced synthetic high-frequency content up to fullband that wasn't in the original signal.

When I was feeding Opus at quite low bitrate with lowpassed signal, Opus treated high frequencies badly. If the lowpass frequency didn't match Opus encoding bands' transition, the encoder produced synthetic high-frequency content up to fullband that wasn't in the original signal.

Quote from: ziemek.z on 2017-02-07 07:40:45

When I was feeding Opus at quite low bitrate with lowpassed signal, Opus treated high frequencies badly. If the lowpass frequency didn't match Opus encoding bands' transition, the encoder produced synthetic high-frequency content up to fullband that wasn't in the original signal.

I can confirm something like this. Here is http://storage8.static.itmages.com/i/17/0207/h_1486492111_5644539_531084a2c8.png the spectrogram of an opus-file at 70 kbps VBR, encoded in foobar2000 with sox-lowpass filter at 16 kHz http://storage8.static.itmages.com/i/17/0207/h_1486492496_8085833_5d1bb7ef27.png
Some kind of high freq "artifacts" are clearly seen on the spectrogram. And interestingly, they disappear (or become hardly noticeable) on upper bitrates (180 or smth.).

Quote from: ziemek.z on 2017-02-07 07:40:45

When I was feeding Opus at quite low bitrate with lowpassed signal, Opus treated high frequencies badly. If the lowpass frequency didn't match Opus encoding bands' transition, the encoder produced synthetic high-frequency content up to fullband that wasn't in the original signal.

Do you have an example signal (with command-line option) that demonstrates that problem?

Yes, this is an example from my New Matura Success Upper Intermediate CD rip. Original is MP3@128kbps, lowpassed at 16 kHz, along with a transcode to Opus@14kbps VBR, encoded with Opus 1.2 beta and corresponding Speks.

Yes I have a very similar experience. I thought it's related to some kind of clipping or resampling, but then it comes from the nature of the codec.

So Opus uses an artifical band above 15.8kHz (according to my spectrum; at least on mid and low range bitrates), and if source is lowpassed at 16kHz it puts some kind of signal derived from that tiny 200Hz signal portion up to 20kHz. This is possibly masked by the <15.8kHz content but I don't like it. The best would be to just say, don't encode anything in that band, but there's no such option at the moment. (i'm encoding the same stream for mobile and the higher bitrate for headphone and home listening) I would like to keep audio up to 16kHz for the higher bitrate setting. So it's a bit of a give or take.

Though i have to say I can't seem to hear any audible issues with the 72kbps stream encoded from the lowpassed signal, so it's probably a matter of paranoia

Well one obvjous issue in the example you posted is the very strong 16 kHz tone in the original. Not sure where that's from, but Opus is definitely not going to like it when all you give it is 14 kb/s. At that rate, there's no way you're going to get good quality music, no matter how you filter.

So Opus uses an artifical band above 15.8kHz (according to my spectrum; at least on mid and low range bitrates), and if source is lowpassed at 16kHz it puts some kind of signal derived from that tiny 200Hz signal portion up to 20kHz. This is possibly masked by the <15.8kHz content but I don't like it. The best would be to just say, don't encode anything in that band, but there's no such option at the moment. (i'm encoding the same stream for mobile and the higher bitrate for headphone and home listening) I would like to keep audio up to 16kHz for the higher bitrate setting. So it's a bit of a give or take.

If you can find a signal where the Opus file *sounds* noticeably better when filtering the original at (e.g.) 15 kHz than at 16 kHz, then please post the original and I'll have a look at it.

foo_abx 2.0.2 report
foobar2000 v1.3.13
2017-02-07 23:32:09

File A: New Matura Success UI Wbk T05-no-sox-lowpass.opus
SHA1: 1cd4d01806184957a8d2b59c2b0bf049c26e981e
File B: New Matura Success UI Wbk T05-15kHz-sox-lowpass.opus
SHA1: d46ff697a0877c17991e96ad4bc24ea55f076e6d

Output:
DS : Primary Sound Driver
Crossfading: NO

23:32:09 : Test started.
23:32:22 : 01/01
23:32:29 : 02/02
23:32:39 : 03/03
23:32:48 : 04/04
23:32:57 : 05/05
23:33:10 : 06/06
23:33:25 : 07/07
23:33:43 : 08/08
23:33:49 : 09/09
23:33:51 : 10/10
23:33:56 : 11/11
23:33:58 : 12/12
23:34:01 : 13/13
23:34:05 : 14/14
23:34:09 : 15/15
23:34:09 : Test finished.

 ---------- 
Total: 15/15
Probability that you were guessing: 0.0%

 -- signature -- 
cc67e7dae19dca11e46b895dacbd86db278044b3

Same recording, but with sinc-lowpass@15kHz applied with SoX. Sounds better without these ugly high-frequency artifacts
And... ABX result (I don't want to violate TOS#8):

Quote from: ziemek.z on 2017-02-07 22:33:32

Same recording, but with sinc-lowpass@15kHz applied with SoX. Sounds better without these ugly high-frequency artifacts
And... ABX result (I don't want to violate TOS#8):

Actually, can you find an example without that strange HF tone? Otherwise it's really hard to know whether the difference is from the "made up" 16-20 kHz band or just from getting rid of that tone (which is also going to cause problems at low rate no matter what). In general, if I'm going to spend time getting rid of an artefact it's both more useful and easier when it's something relatively common (e.g. some mainstream music filtered in condition X) than the combination of a strange signal (e.g. this tone around 16 kHz) in a strange (e.g. this low-pass) condition.

I found some Creative Commons Music in MP3@128kbps. I reencoded it to Opus@48kbps. Same thing happens, but it's not annoying due to higher bitrate used.

You mean that the MP3 you're including is the input signal you're giving to the encoder? Also, what do you mean exactly by "it's not annoying due to higher bitrate used"? As I pointed out earlier, what I was looking for is samples for which applying a low-pass at a lower frequency (e.g. 15 kHz) ends up sounding better with Opus than applying a low pass at a higher frequency (e.g. 16 kHz) or no lowpass at all.

I've got shorter sample without strong HF tone. And... I couldn't find any difference between normally encoded Opus and encoded after lowpassing even when I crossfaded tracks in foo_abx! Another mystery for me is: why LOWPASSED track is BIGGER than normally encoded? I thought that the lowpassed one should be smaller because of lesser value of data to encode...

Quote from: ziemek.z on 2017-02-13 19:49:45

Another mystery for me is: why LOWPASSED track is BIGGER than normally encoded? I thought that the lowpassed one should be smaller because of lesser value of data to encode...

My best guess as to why the band-limited signal encodes to a larger file is that it may cause the encoder to think that the band just before the low-pass is more important than it really is.

If you used an audio editor for lowpass it could also be due to additional dither/truncation noise when converting the lowpass signal back to 16 bit pcm.

I think that dealing with lowpassed signals in opusenc should be improved a bit.

Quote from: ziemek.z on 2017-02-14 09:24:20

I think that dealing with lowpassed signals in opusenc should be improved a bit.

Improved how? You just said you couldn't hear a difference. If there's an actual artefact that should be removed I may be able to do something about it, but if it's only "this spectrogram looks wrong [but it's inaudible]", it's not good reason to change anything.

Quote from: darkbyte on 2017-02-06 19:25:10

I'm evaluating Opus for internet radio purposes (@72kbps and @160kbps) and wondering if feeding a 16kHz lowpassed signal to the encoder would give worse, same or better quality than encoding a full band 20kHz signal? The lowpass actually comes from my multiband compressor and i don't really like to keep content above 16kHz because it seems that >16kHz content may stear the multiband compressor away (some EDM have loud signals in that region).

Low-passing is very unlikely to give better result, but it *probably* won't be much worse either. I've never actually experimented with that, so you should just try and see for yourself.