-
WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!
Opus codec and why its magic
- Thread starter yuki980e
- Start date
BillG
Major Contributor
Our ears are rather insensitive, but a spectral analysis will give one better insights...
While Opus is certainly extremely good, grouping AAC/Vorbis and MP3 is a grave mistake. A bit like saying that MPEG-2 and MPEG-4 AVC are the same.
How? Lossy compression is all about psychoacoustic tuning. Same way we can't really measure video encoder performances without looking, for now; yes, even with VMAF getting better.Our ears are rather insensitive, but a spectral analysis will give one better insights...
BillG
Major Contributor
Last edited:
Do you not know what a spectral analysis is?
I think he is right, spectral analysis won't really tell you how good lossy codec is, you need to listen to them to tell that.
BillG
Major Contributor
Well, you can think whatever you like, but that doesn't mean that's what actually happens when these things are developed and evaluated: a spectral analysis would very much be part of that process, along with ABX testing...
Well, you can think whatever you like, but that doesn't mean that's what actually happens when these things are developed and evaluated: a spectral analysis would very much be part of that process, along with ABX testing...
Don't be so sure - with lossy codecs it's mostly about dropping parts of the signal that are masked and that happens dynamically meaning not always will the same frequency be cut and spectral analysis won't show that. 2 lossy codecs can have very similar spectrum graph yet they will sound completely different. The point is that when looking at spectral analysis you can't even begin to guess about the true strategy of a lossy codec.
Of course it does.
Here is, for example, a spectrum of 500Hz-spaced multitone signal at –10dBFS of AAC codec at rate of 320kbps.
Same thing for MP3 at rate of 320kbps.
So, tell me, @BillG, what can you make out of it?
dc655321
Major Contributor
- Joined
- Mar 4, 2018
- Messages
- 1,597
- Likes
- 2,236
That's a good point - looking at the spectra of the total signal at once may not be terribly revealing.
More useful would be to difference the files in the time domain - lossless minus lossy - then create a spectrogram of that result.
Maybe an experiment for a lazy weekend (i.e. not this one
).EDIT: oops, meant "lossy minus lossy".
That's a good point - looking at the spectra of the total signal at once may not be terribly revealing.
More useful would be to difference the files in the time domain - lossless minus lossy - then create a spectrogram of that result.
Maybe an experiment for a lazy weekend (i.e. not this one
EDIT: oops, meant "lossy minus lossy".
Good idea. Although looking at these 2 spectrums I suspect that even that one wouldn't be able to tell much about differences. I never did ABX comparison between mp3 and AAC but I suspect the difference would be far from obvious.
Depends on the bit rate and the material. At 128 kbps or lower, the differences can be quite apparent. If you want to compare lossy codecs, you need to push them until they start falling apart. That's where the differences will reveal themselves.
While that is true I think that 320kbps (and eventually 256kbps) are most frequently in use and differences there are probably quite small.
Are there tests that show opus becomes transparent at a bitrate, if it's better than aac at high rates, not just low it might manage it.
Is there a good reason it's not being adopted more, other than inertia.
Is there a good reason it's not being adopted more, other than inertia.
There's no shortcuts when evaluating perceptual codecs. Those who have worked on codecs all wish there was, but there's not. Their strategy is to hide noise where it doesn't matter, both in time and frequency domain. So looking at the difference in time or frequency is interesting, but won't inform about perceived quality. The open-source codecs tend to develop *and* evaluate/compare using a psychoacoustic model (PEAQ, etc.). That's cheap and worthless, it's just a fancier way of comparing differences. Modern codecs are at least as sophisticated as the perceptual models in these evaluation methods, so not the right tool for the job.
The only way to evaluate these is by blind listening tests (MUSHRA, ABX, etc) with enough listeners. And with high rates, the first step is to find critical material (stuff that is hard to encode). Examples that come up a lot are applaud (try stereo too), castanets, speech (German is tough to encode for some reason), bagpipes and other tonal instruments. These stress either the time or frequency behavior of a codec.
Have fun, but accept that this is hard work.
The only way to evaluate these is by blind listening tests (MUSHRA, ABX, etc) with enough listeners. And with high rates, the first step is to find critical material (stuff that is hard to encode). Examples that come up a lot are applaud (try stereo too), castanets, speech (German is tough to encode for some reason), bagpipes and other tonal instruments. These stress either the time or frequency behavior of a codec.
Have fun, but accept that this is hard work.
Except that ABC/hr is better than MUSHRA, please, all of that ^^^^
We can make measures of video encoder performance (not ignoring the requirement for controlled subjective analysis)
PSNR as an example.
Last edited:
I'm in a position where I can't comment exactly, but many codecs have a basic design that force a limit to improvement at some given bit rate.
Yeah, I can calculate PSNR, SNR, SegSNR, and lots of other SNR's on audio too.
But what do they tell us?
There is a world of difference between a video codec whose error pattern is, oh, say +1+1+1+1-1-1-1-1 and one that is just +1 or -1, but the PSNR remains the same.
YES?
Similar threads
- Poll
- Poll
