i'd like to add that i organized a test few years ago and not only listeners were absolutely unable to spot the differences between CD and 24/96, but they were struggling to find the MP3's and were completely unable to find the AAC versions as well...
Compressed Lossy AAC files (the standard iTunes) are virtually indistinguishable from any other so-called high quality versions, in a controlled blind testing environment.
Compressed Lossy AAC files (the standard iTunes) are virtually indistinguishable from any other so-called high quality versions, in a controlled blind testing environment.
Are you referencing the Meyer/Moran 2004 study?
This is really old news. I mean like really really old news.
It's been discussed to death in many places.
^ Yes, but amusing in light of Jon's other "best DAC" thread. Whether deliberate or incidental irony is at hand, I'm enjoying it.
Opening RDF's article now--will give it a proper read.
i don't see the contradiction: a DAC can perform well (or bad) even on a lossy AAC or MP3, even though its all prepped-up for sky-high resolution..
Last edited:
? It doesn't mention MP3 in the abstract, though I haven't read the article yet.
This excerpt tells me it was a null test that the authors tortured the data with statistical software until it confessed. (I hope this quote runs no copyright foul--if so, please axe it immediately!) N=3 subgroup analysis on a N=16 set? Eeeesh.
Their actual testing methodologies did not in any way give me pause--they seemed quite solid. Likewise, if we simply take their subgroup analysis, the 3 individuals could tell the downsampling step, but not between native 44k and native 88k.
All listeners claimed the test was extremely difficult, and outside of the abstract, the claims made within the paper were a bit more muted.
Please note, I haven't gone through the 2004 article with the same rigor, so pretty please do not assume that my laziness implies tacit approval.
It's possible, further trials with their subgroup would sort it out. I read another paper ages ago describing listening trials between two ri-rez formats that similarly suggested an individual demonstrated standout ability to correctly identify between the two. That also wasn't followed up.
The only discussion on this paper I could find online were forum chat, blogs, etc., nothing substantial but no critical analysis either. Yours is the most in-depth so far. However I've also read tests with null results in which no one bothered to test if the playback devices were actually capable of approaching 24 bit performance. A highly publicized trial consequently used as one of their three players a unit that in independent testing barely eked out 17.5 bit playback.
Thanks for the analysis.
You're welcome. Apologies that my analysis is the most in depth! 
The DAC used was a RME Fireface 800. The microphones were fed from a preamp/splitter into two RME Micstacy's in parallel, one recording 88/24 and the other 44/24. No mention of whether the data was downsampled to 16 bits, but I'm going to assume that it remained 24 bits throughout its chain. RME's stuff is generally excellent, but we should not rule out that the signal chain limited the test materially.
As far as my point about a null test--I will dig my heels in on that assertion. I will definitely allow for the hypothesis that 44k vs 88k is audible, but this test in specific did not prove (nor disprove) that. And we can also, with pretty decent confidence, say that any ostensible differences are going to be small. Very small.
Doing further testing with just the 3 individuals who could tell the difference (note-not correctly match A to A or B to B, but could discern A from B, and barely), would be to ignore the other 13. One could argue that we should set our standards above the "most sensitive" listener, as that covers all bases. I'm willing to concede that.
With that in mind, however, the most-but-barely significant discernment was downsampled 88.2->44.1k vs 88.2K native and downsampled 88.2->44.1 vs 44.1k native, which leaves you wondering how "good" the proprietary resampling algorithm is. 88k native vs 44k native only showed to be better than p > .05 when they stripped down to 1 of the 5 recordings (within the 3 individuals only, to boot!), as the other 4 recordings were not significant--that is seriously torturing the data to get it to confess.
The DAC used was a RME Fireface 800. The microphones were fed from a preamp/splitter into two RME Micstacy's in parallel, one recording 88/24 and the other 44/24. No mention of whether the data was downsampled to 16 bits, but I'm going to assume that it remained 24 bits throughout its chain. RME's stuff is generally excellent, but we should not rule out that the signal chain limited the test materially.
As far as my point about a null test--I will dig my heels in on that assertion. I will definitely allow for the hypothesis that 44k vs 88k is audible, but this test in specific did not prove (nor disprove) that. And we can also, with pretty decent confidence, say that any ostensible differences are going to be small. Very small.
Doing further testing with just the 3 individuals who could tell the difference (note-not correctly match A to A or B to B, but could discern A from B, and barely), would be to ignore the other 13. One could argue that we should set our standards above the "most sensitive" listener, as that covers all bases. I'm willing to concede that.
With that in mind, however, the most-but-barely significant discernment was downsampled 88.2->44.1k vs 88.2K native and downsampled 88.2->44.1 vs 44.1k native, which leaves you wondering how "good" the proprietary resampling algorithm is. 88k native vs 44k native only showed to be better than p > .05 when they stripped down to 1 of the 5 recordings (within the 3 individuals only, to boot!), as the other 4 recordings were not significant--that is seriously torturing the data to get it to confess.
Ha ha, some how that does not surprise me. I catch a lot of friends listening to those awful compressed satellite radio feeds without even noticing. Yikes!
I've also attended several casual listening sessions where people preferred the 192K (or better) MP3 feed over non-lossy. That's understandable, high bitrate MP3 usually sounds warmer and more forward than its original file. Easy to prefer that.
I catch a lot of friends listening to those awful compressed satellite radio feeds without even noticing. Yikes!I've also attended several casual listening sessions where people preferred the 192K (or better) MP3 feed over non-lossy. That's understandable, high bitrate MP3 usually sounds warmer and more forward than its original file. Easy to prefer that.
Can't quibble with that position. Personally the music I find easiest to enjoy and spend the most time with are 24-bit live recordings from the Internet Archive. Granted, it might still be the case that anyone willing to take the time and effort to haul DPA mics and 24/96 equipment to live concerts will also be the most likely to expend the effort capturing a good recording, skewing the perspective.
What's possible is more interesting to me than what happens centre of the Bell curve. I also don't discount the possibility of learning differences. When the first ISO codec MP3s became available online there were some ferocious arguments defending their 128 kilobit perfection. Hard to believe today but it took time for many listeners to hear the nasty swirl of hash and distortion. Somewhere I have a training CD - maybe AES - to teach hearing the effects of data compression. The same might happen with 44/16 as listeners get more exposure to true HD source material.
Ultimately though none of this will matter, storage and processing power are becoming so cheap that compressing to MP3 or mastering to 44/16 will be a complete waste time no matter what the general perspective is on HD audio.
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.