Skip to main content

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: Opus killer sample at 256kbps (Read 3922 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

Re: Opus killer sample at 256kbps

Reply #25
I tried wavpack with this HF lab sample. 320 is easy 8/8. more forward aggressive tonality.
384 is very close 7/8, slightly different presentation.  448 failed abx
settings;  -hhx6s.75  -b3.63,  -b4.35,  -b5.08

wavpack -b3.63hhcs.5

Re: Opus killer sample at 256kbps

Reply #26
While testing for opus transparency i found that it does not do well when there is a lot of high frequency noise,
You can get an improvement with --framesize 10.
------
Chirp0to24kHzPlusWhiteNoise - same chirping on test file (15kHz region);
Chirp0to24kHz - from 12kHz onwards it experiences convulsions (the straight line increases with jerks. MP3 and AAC go smoothly).
I watch through Spectrum from Foobar.
Test files > Chirp0to24kHz.zip (10.51 MB)










Re: Opus killer sample at 256kbps

Reply #27
Developers may not hear and continue to ignore problems at high frequencies ::)

Re: Opus killer sample at 256kbps

Reply #28
If we think this behaviour is acceptable we need a huge disclaimer on the OPUS codec.

“NOT SUITABLE FOR VIDEO OR VIDEO GAMES. TRANSPARENCY CAN ONLY BE ACHIEVED WITH SPEACH AND “NORMAL” MUSIC”.

Re: Opus killer sample at 256kbps

Reply #29
screens added, all from pressed, commercial CDs.
Please name these CDs and share audio samples of these tracks.

Why? They’re not the result of hundreds of hours of research, it’s just a collection of tracks that I thought had “sharp” sounds.

One of them is from the Jimmy Edgar EP “Make, Model, Bounce”.

You never heard of Pan Sonic?

Re: Opus killer sample at 256kbps

Reply #30
The output from the codec is audibly different from the input. You can argue that it’s not a common occurrence, but it remains a problem.

I appreciate the fact that you're completely ignoring my point and only focusing on points you can come up with valid responses to, but let me reiterate:
speaking of screens and things we actually do perceive with eyes, imagine this approach to judging lossy video encoders. Arguably, those are even easier to break and with more realistic material than that provided in the OP: just some confetti or a moderate snowfall is enough to bring any lossy video encoder to its knees and render its frames to look like pixel soup. And yet, we accept this as reasonable and don't mind.

I’m not ignoring it. It’s a terrible analogy. Lossy video is almost never “transparent” to the master. In fact we generally don’t even shoot in lossless formats for video any more.

Re: Opus killer sample at 256kbps

Reply #31
It’s a terrible analogy. Lossy video is almost never “transparent” to the master.
I often see the video people use the phrase "lossless" to mean perceptually indistinguishable from, that is what we call "transparent". In that terminology, they don't use the word "lossy" if it is transparent.

They also have a "slightly different" attitude: if you want to distinguish by eye, you can zoom in by insane factors. They assume you don't. Wouldn't work against audiophool users.

In fact we generally don’t even shoot in lossless formats for video any more.
Now back to "our" terminology and yes you are right - but inconsistent with previous sentence. Shooting and storing directly to a mathematically lossy format and then re-encoding to the same - mathematically lossy, but, if high enough resolution, it is transparent.



Last two months' worth of foobar2000.org ad revenue has been donated to support war refugees from Ukraine: https://www.foobar2000.org/

Re: Opus killer sample at 256kbps

Reply #32
Lossy formats are all about sweet spot and trade off, among bitrate, processing power, latency and quality. For video games, CRI ADX, a proprietary ADPCM format is very popular. There were also XA-ADPCM for Playstation games. These formats don't have good bitrate/quality ratio, but they have low processing power requirement and latency.

Other formats like aac, mp3 and vorbis exhibit weaknesses in other aspects as well, some of them are discussed in the Listening Tests sub-forum for example.

Re: Opus killer sample at 256kbps

Reply #33
Lossy video is almost never “transparent” to the master.

Well hold on just a second there, buckaroo. Where is this claim coming from? What is "almost never"?

I can't comment on this claim with anything other than a retort that it is simply blatantly incorrect. If video is encoded with high enough bitrate, most video content will be transparent - with the exception of the aforementioned content of confetti and moderate (or more severe) snowfall, which essentially won't be encoded transparently no matter how high a bitrate is used. And that is exactly the same situation to what we have here with Opus - except one can argue that the particular content that breaks Opus is more artificial than confetti or snowfall.

This isn't a "terrible analogy", it's an almost perfect analogy - you just don't like it.

They also have a "slightly different" attitude: if you want to distinguish by eye, you can zoom in by insane factors. They assume you don't.

Not only that, but you also aren't supposed to be staring at a single individual frozen frame of a video - you're only supposed to see an individual frame for a couple dozen microseconds, and video encoders actually use this fact as well to spend less data on very quickly moving/changing areas of a frame. If you pause and stare at an individual frame, you're "cheating". In other words, if a video encoder failed to achieve visual transparency when looking at an individual frame, that does not also mean that it has failed to achieve visual transparency when the video is actually watched at its nominal framerate.

Why? They’re not the result of hundreds of hours of research, it’s just a collection of tracks that I thought had “sharp” sounds.

One of them is from the Jimmy Edgar EP “Make, Model, Bounce”.

By the way, did you actually encode these tracks using Opus and verify that you can ABX them? Or did your entire argument stop at proclaiming that their spectra look "similar enough" to the problematic sample?

For that matter - I see you also haven't ABXed the sample provided in the OP - you just jumped on the bandwagon without even checking that you can hear a difference.

Re: Opus killer sample at 256kbps

Reply #34
For anyone interested in further testing, I provide more encodes of the sample at varying frame sizes (frame size indicated in the filename). All provided encodes are 256 kbps VBR.

Re: Opus killer sample at 256kbps

Reply #35
Why? They’re not the result of hundreds of hours of research, it’s just a collection of tracks that I thought had “sharp” sounds.
For testing purposes, obviously. I don't understand how this is music in the first screenshot.
So, can you upload the corresponding audio samples please ?

Re: Opus killer sample at 256kbps

Reply #36
The whole purpose of lossy codecs is to range in functionality from "not terribly annoying" to "audibly transparent to the listener". The threshold depends on how much you're willing to put up with, and how much bits you're willing to throw at the encoder. Clearly, in this case, transparency depends on the listener, and if it's not achieving transparency even at high settings, something needs to be looked into.

Re: Opus killer sample at 256kbps

Reply #37
The whole purpose of lossy codecs is to range in functionality from "not terribly annoying" to "audibly transparent to the listener".

Honestly, I feel like this perfectly summarizes why non-transparency in this particular case is such a non-issue to me.

Well, first of all, it's noise. It's not white noise - rather it's particularly coloured noise - but it's noise nevertheless. And there's only so much you can do when encoding pure noise, even when encoding lossily. To me, the original sample and the Opus encode both equally adequately sound like particularly coloured noise, even though they sound (slightly) different.

But onto the part that relates to what you posted. Which is - even though, from what little I listened to the original sample and the Opus encode, to me it seemed like I could hear a slight difference, both samples were too PAINFUL for me to actually perform a proper ABX. Personally, I think that in itself is relevant to judging how detrimental a failure to achieve transparency is for a particular sample. When lossy audio encoders are tested, people are asked to rank samples on a scale which includes options such as "noticeable but not annoying", "annoying", etc. In this case, I would personally rate the original sample itself at the absolute bottom of that scale, so from that perspective, what even is there for a lossy encoder to ruin further anyway?!

To me, both the FLAC and the Opus samples make my ears bleed - it's just that they do it while sounding slightly differently. Does one sound better/worse than the other? To me personally, no.

If - somehow - a piece of music I was listening to included a sample such as this and the Opus encode failed to achieve transparency in the same manner in which it failed to achieve it here, would I notice it and would it bother me? Even if I did notice it, there is absolutely no way that it would bother me more than the original sound itself bothers my ears in the first place.

If something like this happens to a sample for which I think the Opus encode actually makes the sound worse, then I'll be bothered.

Re: Opus killer sample at 256kbps

Reply #38
None of the samples were PAINFUL for me to actually perform a proper ABX.

Everyone is talking about "music" or whatever. Lossy audio encoding is not supposed to be used exclusively in music.

We have lots of examples of "noise" that are transparently coded at these bitrates. I assume this is by design and not just an accident.

I can understand the limitations of lossy encoding but spinning it to be a good thing outside of niche scenarios (eg. using artefacts on music) really takes the cake.





Re: Opus killer sample at 256kbps

Reply #39
spinning it to be a good thing

This is a thing I very explicitly did not do.

What I said is that I don't see the failure to encode this particular sample transparently as detrimental, because I personally am primarily interested in whether a lossy encoder degrades the sound quality, and to me both the original sample and the encode sound equally bad. Nowhere in this do I claim that the failure to encode this sample transparently is "a good thing", just that I personally don't find it problematic.

I'm not spinning anything, but you are - you're spinning my words into something I never said.

Re: Opus killer sample at 256kbps

Reply #40
This is maybe a goes-without-saying to some, but likely not to everyone:

Lossless encoding depends on the information content of the signal. Harder to compress is harder to compress.
Lossy encoding depends on the (most) useful content of the signal, but not directly - it depends on a model for ordering the "(most) useful".

White noise is hard to compress losslessly, but in principle easy lossy - just replace it with a "play noise!" instruction.

A compressor should minimize overall badness for given total size. Saying that a lossy compressor shall not judge which music shall have artifacts, gets two things wrong:
1: a lossy compressor shall obviously compress silence with no artifacts because that is easy - even if that means that difficult music will get treated worse.
2: a model for hearing is a model. All models are wrong. Some are useful.

Saying the opposite also gets something obviously wrong: if those who build and tune the psy model didn't imagine we would listen to these or those signals, then well, it may have failed out-of-sample testing. Revise.

Saying that "minimizing badness entails that ..." often makes the mistake that there is one nature-given criterion with no subjectivity. It isn't. Lossy compression at low bitrates is about subjective badness. Hopefully my subjective badness when I am listening (and yours when you are).

And pulling the "so you are saying we cannot even assess lossy ..." forget that thing. Even in situation where there is no subjective agreement on what is best, there could very well be a ton of consensus about wrongs. Yes we can say something about the wrongs of [fill in nineties encoder at suitable bitrate] and whether we have improved.
Recommended: Isaac Asimov's "The Relativity of Wrong". The ... uh, the title track, in musical terms. Google for (errhm something something TOS9).
Last two months' worth of foobar2000.org ad revenue has been donated to support war refugees from Ukraine: https://www.foobar2000.org/

Re: Opus killer sample at 256kbps

Reply #41
Lossless audio encoding is shortsighted - if you feed it with a signal bit-for-bit repeating every minute several times, it won't be very effective. If you feed 7z or rar with it, given large enough dictionary, they will be several times more effective. It's sometimes the case with synthesized/emulated beeper/pc-speaker or otherwise oldschool chiptune music. That's the cost of specialisation.

Lossy encoders are specialised/tuned for real world TYPICAL audio signals. There ARE edge cases where they will fail either by lack of CBR/ABR transparency or by VBR bitrate bloat. Is it a problem if it happens with typical audio signals? Yes. Is it a fail if it happens with 0,01% of artificial signals? No. Should it be looked into? Yes. Should it be tuned if there's a risk of affecting typical audio encodings? No.
Lets take a screenshot of this page and save as 80% quality JPEG. Artifacts are obvious. Is JPEG a fail and should it be tuned? No, 80% quality is perfectly fine for 99% of photos.

And indeed that artificial example from this thread isn't very annoying to me, barely noticeable, far from emotions it aroused. But I always was on the lower side of bitrates (most of my collection is MPC --radio).
I'm guessing now but it probably is possible to filter the original file the way an audio engineer would do - so it still sounds the same but doesn't have that absurd ultra noise, and I guess Opus would encode it well then (seems like it allocates too much bits for that useless freqs which should be lowpassed whatsoever).

Re: Opus killer sample at 256kbps

Reply #42
Is it a problem if it happens with typical audio signals? Yes. Is it a fail if it happens with 0,01% of artificial signals? No. Should it be looked into? Yes. Should it be tuned if there's a risk of affecting typical audio encodings? No.

100% agree.

Re: Opus killer sample at 256kbps

Reply #43
Me too. As a codec developer, I actually looked into this by encoding at different bitrates and can't find any obvious problem here. The 14-kHz spectral content is simply at the threshold of being zeroed out during encoding. Anybody worried by such behavior should either increase the bitrate (here, to e.g. 320 kbps) or use a different codec (i.e., a newer one).

Chris

P.S.: I can't hear the artifact myself, at least not at loudness levels which don't hurt my ears.
If I don't reply to your reply, it means I agree with you.

Re: Opus killer sample at 256kbps

Reply #44
If you feed 7z or rar with it, given large enough dictionary, they will be several times more effective. It's sometimes the case with synthesized/emulated beeper/pc-speaker or otherwise oldschool chiptune music.
Synthesized music will rarely have exactly repeating sequences because its samples may start at an offset in time that is not an integer multiple of a sampling interval, such as if the music has been rendered at an odd internal rate, or if it has been captured from a physical device. 7-Zip by default has a delta filter for WAV files (only), which has to be defeated to take most advantage of repeating patterns. Other compressors also have multimedia mode for sound and graphics, similar to how a lossless audio codec works.

I agree that it is not worth optimizing the codec for artificial noise. It is well preserved in this example. If it was used as an effect in a real audio file, to simulate a raging water or an electronic device like the HBO logo, I would get the same impression from it after compression.

What I initially thought to be artifacts, three clicks (at 0.57, 3.20, 6.1 seconds) and three tones are also present in the original. There seems to be a real sound mixed under the noise, or a random generator that is not truly such. To hear the high frequency portion of the spectrum, I have to turn up the volume, and then I only hear a sharp pulse when starting playback, as my ears rapidly become insensitive.