I'll add to this that "high bitrate" for Opus can reasonably be considered as 128kbps or above, generally 160 or 192kbps are comfortably beyond what is needed to achieve audible transparency.
For anyone who's used to MP3 and just doing everything at -V0 or 320kbps, it's extremely impressive to hear how good Opus is, and how low bit rates you can get away with. Even at 24kbps it's serviceable for music, it sounds no worse than a slightly worn cassette tape.
It's a seriously impressive codec, both for speech and musical content.
24kpbs is perfectly adequate for voice/podcasts/audiobooks
This is such a key point: MP3 has known weaknesses with certain common sound profiles (e.g. sharp transitions such as a cymbal or snare drum) which aren't really fixable. For me the threshold was AAC 256kpbs where I stopped noticing artifacting.
On the whole your comment is great, but I think this point might be revealing of either your age or the last time your looked at mp3 encoders. LAME has been great for at least a decade now. Other than (possibly) extremely rare killer samples, you're unlikely to be able to ABX a LAME encoded track at 256 Kbps. I certainly can't, and I have pretty decent gear and have been listening to (and working on) music for a very long time.
Back in the 90s, all mp3 encoders were shit. Xiph's Monty, in the post you link, mentions being able to distinguish between them using only their results as a party trick. Even in the early 2000s you needed 256-320 Kbps to have a shot at transparency. But now, most music is transparent at `lame -V2` settings, which gives approximately 192 Kbps results. I can still ABX a very small number of songs (last I tested), but even these go away by `lame -V0`, which is the highest quality VBR mode. HydrogenAudio, a trustworthy source as far as "audiophile" claims go, says that anything from `-V0 to -V3` should be transparent under most conditions. https://wiki.hydrogenaud.io/index.php/LAME
AAC and Vorbis are certainly better codecs, achieving transparency most of the time at closer to 160 Kbps. Opus is even better. I've never been able to ABX a 128 Kbps track, and music sounds great even at 96 Kbps (for stereo!). There are supposed to be killer samples for all four, forcing them to require a higher bitrate to achieve transparency, but interest in finding these seems to have dropped.
But storing music over 48kHz and 16bit for comsumption is really just waste of storage space and pure snake oil. 16 bit give a dynamic range of 96dB and I'm not aware of any recording outside of experiments that take advantage of it. Even the most expensive speaker / headphones will distort terribly if you play them on the loudness level where this dynamic range would matter...
And most "high quality" recordings with a high frequency range (96kHz+) only really add ultrasonic sound and mostly noise artefacts that don't correlate to the music recorded (and that no speaker / headphones can even reasonably, without high distortions, reproduce, if they even pass the LPF of the reproduction chain).
So yeah, I also think that 48/16 is all we need for optimal consumption. If the format is lossless, it's also nice for archival reasons.
http://www.lavryengineering.com/pdfs/lavry-white-paper-the_o...
I'm personally baffled and irritated by this move from Apple. They'll be selling healing crystals next.
With most samples, yes. The issue, as I said, is that psychoacoustic codecs are basically heuristic by definition, and so you can pretty much always find counterexamples. It's possible to say 320kbps MP3 is indistinguishable from uncompressed for most samples; it is not possible to say it is indistinguishable for all samples, present and future.
It's also the case that lossy encoding is a bad idea if you're going to be further processing the audio in any way; repeated transcodes definitely start bringing out the artifacts more. So lossless music is always something worth having as an option.
Now, when you start talking about "high-resolution" 96/192k 24b stuff... yeah, that's just as good as healing crystals.
Are there likely to be real-world cases where this happens - and the effect is detrimental enough to actually matter to anyone?
Surely we're debating "when is it good enough for anyone sane and reasonable?" not "when could a hypothetical oracle be fooled 100%"?
And as I said, really, the main reason to avoid lossy compression at this point is due to generational loss and post processing. Lossy compression should be considered a final processing step - what you do to store music that is then going to be delivered directly to a listener, unaltered. If you're going to do anything else, you'd do much better with a lossless version.
That is not to say, of course, that if all you have is a lossy version, it is a major problem :-)
For the record, my lossy format of choice for e.g. putting stuff on my phone is 96kbps Opus. Even that much is excellent for casual listening. But I much prefer to keep lossless FLACs as my primary archival storage - not just because that way I can take advantage of better compression formats as they come (e.g. how I moved from ~130k Vorbis to ~96k Opus), but also because compression makes a massive difference with certain kinds of processing and editing which I sometimes enjoy doing. All the psychoacoustics go out the window if you start doing things like subtracting instrumental mixes from full mixes to get vocal tracks out.
Ultimately though, lossy compression is always program-dependent ("program" here means the song, not the app). The encoders use psychoacoustic algorithms, and there is no way to prove they will perform well on all samples. So you will always be able to find corner cases where an encoder or format does worse and results in more audible artifacts; some of these might be encoder bugs, some might not. Therefore, it is fair to say that you might not want to rely on lossy compression for archival. But it is perfectly reasonable to use it to put music on devices to actually listen to.
You say good mastering makes it easier to notice encoding artifacts, but actually, bad mastering can do that too. Over-compressed stuff (in dynamic range) with inter-sample peaks above 0dB can clip and distort after getting run through a lossy codec.
As for lossless formats, though, anything above 48kHz 16 bits is complete nonsense for final delivery to consumers. All the "high resolution" stuff with higher sample rates and bit depths is just marketing bullshit, and this has been repeatedly demonstrated in trials (the well-designed ones anyway; there are plenty of terrible ones - no, comparing the 96kHz download to the 48kHz download of a song is not how you test this properly).
Bit depths above 16 are useful during production (because quantization noise accumulates and is boosted by things like dynamic range compression); in practice you want to record at 24 bits for headroom reasons, and process in 32-bit float because there's no reason not to with modern computers (and many practical advantages, e.g. ~infinite dynamic range). Sample rates above 48kHz are less so; the main reason to use them is to avoid aliasing artifacts, but it is usually much more effective to do that with well-designed DSP algorithms that include internal oversampling, as opposed to just doing everything at a higher sample rate. Unless you're doing extreme pitch shifting; then higher sample rates may make sense.
Videos everyone interested in audio production should watch:
https://xiph.org/video/vid2.shtml https://www.youtube.com/watch?v=-jCwIsT0X8M
The little tidbit at the end of the second video about how 48kHz is, in some respects, twice as good as 44.1kHz, is quite neat and not something I'd realized before.