Are inter-sample peaks REALLY audible ? (Hint: YES)
HTML-код
- Опубликовано: 8 сен 2024
- More and more people are saying that inter-sample peaks (ISPs) are only a theoretical concern, and we'll never hear them. Some have even gone as far as to call them a hoax.
But broadcast and streaming guidelines say we should always avoid them.
So are they really audible ?
In this video I explain what ISPs are, and test using a real-world example to see if we really can hear them, or not. See what you think !
I think I finally understand why some music that gets uplouded to youtube appeared to me as if the "high end" was shaven off a bit. It was probably the ISPs that were cut off. We did a test a while back and uplouded 2 of the same masters just that one was -1.0 dbTP the other -0.2dBTP and the version with -1.0 dBTP was significantly better sounding (also had more transient punch). I suspect now that since the track wasn't ISP limited, that this was probably one of the things I was hearing. Thank you so much for your videos and in also in general! Your information has helped me numerous times and I can always count on you providing clear and concise answers!
The worst part about this is that even though we know this, I've had to deliver files I know full well will suffer the consequences. - the fact that so many favourite artists/genres/etc are all peaking in this way leads sooooo many artists to insist that their tracks be mastered the same way, no matter what advice or recommendations we can make. Of course it's all relative, but I've delivered files for very very happy customers that will surely have ISPs. We do what we can.
It blows my mind that professional engineers are doing the number 1 thing i thought was mixing and mastering 101: don’t allow anything to go over 0db. Very informative video👍🏼
what are you talking about at all
@@triple_x_r_tard i think its pretty self explanatory what i was talking about. If you're lacking knowledge, thats a you problem.
Thank you for this. A "professional" mastering engineer said in a forum that inter-sample distortion was total bullsh*t and I was an idiot for thinking people can hear it. He actually got quite abusive and personal. But even people who think they can't hear it will admit there's a kind of fuzz if audio with inter-sample distortion goes through data compression (e.g., MP3).
The Anti Streaky
😂
The video is interesting.
I used to have my WIndows set to 100% for a logical reason, one that people have told me in the past: it sends all of the bits to the DAC. Thanks to this video, I might try setting Windows to 50-90%, currently trying 50% first. 50% would mean I have an output exactly at 23 bits in theory (24-bit data divided by 2 means you just drop the last bit off the end. 51-99% is still 24-bit, but the audio won't be adjusted by an integer scaling). It does seem easier on my ears if I set Windows to 50% volume, but that could just be psychological. When I had it at 100%, it sounded more brickwalled! I just tried this today after seeing this video.
I also noticed that Windows had effects on, which made the brickwall sounding worse.
One thing about 32-bit audio: it can go past the usual peak. However, most DACs are 24-bit or 16-bit. Want to see what I mean? Take some audio, intentionally peak it, and then save it as 32-bit floating-point uncompressed. Load it back and then normalize. You often have the original sound data. It would be an advantage to 32-bit DACs, able to handle these peaks without clipping.
Windows internally processes audio at 32-bit if it needs to process it.
Great test, thank you so much for the video!
I could clearly hear the difference, and I surely prefer non-cliped version. I think ISPs should be a thing to be aware of. I personally always use final true-peak limiting on my clients' master files. It happens to be a good practice.
Agreed 🙂
Informative video! The differences are clearly audible (to my ears, anyway) and reinforces that paying attention to detail, critical listening, and vigilance are crucial for delivering outstanding masters. Thanks so much for sharing! -chaz
My pleasure 🙂
Great Video! This was really informative and interesting. And good on you for doing the random listening test!
Thanks !
Love your videos thank you so much for making them!
Thanks ! My pleasure 🙂
Its worth clarifying that each sample has a maximum numerical value due to the bit depth, so a "clipped" sample peak is not playing louder to cause distortion rather its more that the peak is shaved or flattened which means the sample bit value is sent to the device multiple times which causes the distortion.
It hasn't changed my mind because I always leave 1 db of headroom and have done similar tests, but the way you explain this just made is that much clearer to me! Also love that you didn't use the true peak on the SIR that
s also what I do, always use a clean limiter after the SIR clipper just for safety!
Thanks, glad you liked the video 🙂
I always wanted to do a video about that myself! Thanks Ian for saving me time😜🥃 - Brilliantly explained, as always!
My pleasure, thanks !
Great video! Kudos for doing the blind tests. Hofa blind test and Perception AB are an indispensable pair for me doing ozone development!
Thanks Bill ! I wish HOFA had an X feature, I actually find that can help sometimes. And it’s really quite easy to get 4/5 through luck and feel as if something is audible when it isn’t. Still useful, though
Thanks for the very informative video Ian. Much enjoyed. Two things i take away from this video
1) Why in Gods name does anyone „need“ to ride levels so near the ceiling. Just cut the music some slavk everyone!
2) Gee this song sounds so flat and stressful that i dont think i could listen to the whole length .
The drums sound like its a drumcomputer playing. Everything squeezed out. Why oh why…..
and i love the red hot‘s but this style of madtering is such a disaster
As to your question… well it goes hand in hand. The ones that risk I.P are usually things that are mastered superhot anyway so
Usually the ones causing trouble are „usual suspects“ in a way. So i dont like their sound way before i would ask myself „hmm could it have I.P issues?“ :-)
Am interesting test might be to play something back which contains a lot of intersample peaks and record the output to examine what a given DAC is doing with those peaks. Or multiple DACs to see if any handle them in a different fashion. Do any/all of them actually clip at full scale or do any have some headroom to account for this?
My understanding, and what correalates to what I hear, is that an ISP that clips causes a wide-band impulse of intermodulation distortion which renders as a kind of noise. To me what is important is what I call "ISP velocity" or "density" - the number of ISP clips per unit time. The more there are more inter-mod distortion there is. My take on this is that this causes a kind of low level "mud" which can rob PLR by increasing over-all loudness. This robs dynamics and can dull transients. I worked with LVC audio to have his ISP meters show "velocity" or "density", Youlean also shows this with red dots at the top of the timeline. Another nice tool is Apple's RoundTrip but it's AU only. It will actually count the ISP clips using different codecs. In my experiments I have found, for my music, that -16 LUFS gives a good balance of loudness vs threat of ISPs. It does mean the trakcs are lower in level but it minimizes the chance of up-stream clipping because of ISP, which is important to me. But If loud is your thing, just make sure you use a good tool because its not always subjectively easy to hear it. It's one of those digital things that benefits from digital tools to measure it to eliminate it.
Interesting, I tend to agree although personally I'm usually comfortable mastering up to -11 LUFS if the true peaks are at -1
I really enjoyed this video! Great info as always! I always try to keep my recordings as "clean" as possible so this def helps keep a few important things in mind when mastering.
Spotify and Apple Music allow users to leave loudness normalization off, so the streaming services won’t change the loudness of the songs you play. I wonder how many people do that (I leave it off). Given that people leave loudness normalization off, it’s tough to say if the decrease in loudness needed to avoid intersample peaks is worth the reduction (i.e. loudness wars). Maybe another test you could have done is between the no intersample peaks version (which is -3dB quieter) vs. the original file that has intersample peaks (but is 3dB louder). Which one would a listener prefer? Basically, is avoiding intersample peaks worth the drop in loudness? Streaming services could force loudness normalization for every song and that would lead me to avoid intersample peaks instead of pushing my track as loud as it can reasonably go.
Fair question - normalisation is on by default on Apple and Spotify, and on Spotify less than 17% of people disable it.
As far as whether it’s worth the reduction, for me the answer is Yes - in fact I’d turn it around and say is the loudness worth the price you pay in terms of the sound - but many/most people disagree with me…!
The clipped version sounded a little more rock & roll! ........... this was a pretty extreme (distorted) example, the real difficult choices come when the clipped version sounds punchier but not really more distorted, then your finger gets itchy to switch ISP protection off.
Crazy interesting.
on some dacs this is really insane... bad enough intersample clipping causes effect like volume is fading after the clip and then it comes back up... audible... well... very much so when it sounds like somebody turns volume near 0 and back up real fast... :D i dont need to even abx and unsuprisingly when appling limitter the volume rollercoaster effect is removed
Great video and audio examples! I do care about them. Most of the time, clients have preferred nonclipped versions, except for two times that I recall. One where the well-produced, intentionally clipped version actually sounded better than the nonclipped version on Soundcloud's crap 64kbps codec, even when doing test exports with increasingly lower bit-rate codecs. The other time, a client preferred the more clipped version, even though there were audible problems with it, so it is what it is.
The issue I have (and meant to mention in the video) is that ISPs are unpredictable. Some converters will clip, others will reproduce the signal cleanly. So even when people prefer a clipped sound, in my opinion it’s better to incorporate that into the master itself, and not leave the final result to chance.
@@ProductionAdvice Agreed!
So I bounced a track to tape and then went through Mix:analog's AD converter. Nothing then should exceed 0dB in the new digital track. Is it then necessary to turn the song down an entire dB or can I get away with less reduction, say, -0.3?
The problem with this is that your example is really extreme. +2db true peak is quite an insane amount. The average professional loud track (-5 lufs to -4lufs) with clipping only goes up to around +0.5db to +1db. And if you have to try so hard to hear the clipping here, it's definitely a non-issue on those other tracks. So while you have shown that you can hear intersample peaks, it doesn't mean that you should kill your loudness with the recommended -1db to -2db ceiling on your limiter. Overall I think it's not really a big issue.
So summing up to avoid that problem of lifeless masters with lossy streaming services codecs it`s better to leave True Peak under -1dB? Can you please explain the concept of the video briefly in more human language, as I`m even more confused than I was. Thank you
Thanks for the video. How about uploading these two song clips you compared, say to Dropbox or some other site, so we can then download and hear them without RUclips compression?
Yeah, even via RUclips codecs the difference is apparent. Thanks for the great reminder!
Yes, there is an audible difference in your example, albeit subtle. However, IMHO the crux of the discussion *should* be about why to avoid ISPs altogether - which you touch upon towards the end by pointing out that streaming services turn it down further. Unfortunately, in a world where proverbial noise pervades nearly every aspect of our lives, a healthy amount of headroom and dynamic range are both fighting - ahem - an uphill battle. I thought the loudness wars were over?
Sadly not 😕
For me, the kick and snare sound a little more buried in the non-ISP version, but I like the sound of both of them. I really think it's dependent on each track and its arrangement. For example, Madeon's 'Icarus' has more obvious clipping in the verses, but as soon as the drop hits, the drums, synths and FX, that gets introduced in the arrangement mostly masks the clipping.
Fair enough - I really dislike the "flat" quality to the clipped (non-ISP) version, especially given how heavily limited and clipped the original is
Something I don't quite understand is where in the D/A does it clip, is it the analog section or in the actual conversion process? I'd assume the reason some people don't hear it is because certain D/A would have higher than 0dbfs headroom but consumer don't which is why maybe engineers don't hear it when rendering. Obviously the main problem is lossy codecs because that creates digital peaks. I have heard plenty of soundcloud posts of songs which sounds clipped to hell, but then the wav beatport release sounds fine, I assume this is because of intersample peaks or worse just peaks in general.
It depends - some converters reduce the level to add headroom for ISPs (Burl for example, I believe) while others don't, but have enough analogue headroom for it not to be a big issue for straight PCM. Many don't have either.
DSP clipping (like the decoder clipping in this video) is a bigger issue though (IMO) because by the time it hits the converter it's too late, the clipping has already been baked into the fixed-point decoded file
I would like to ask an amateur question from someone who is more interested in quality play back then production. If I have a file that has been mastered with ISPs, and assuming I can hear the degradation that it causes, would I still get that distortion if I set my dac to, say -3dbfs ? Secondly what difference would it make if that file was lossless, presumably the ISPs would not be as noticeable, but would still be audible? Please remember I am an amateur, and this question might not make sense to someone who understands this topic! Thank you
Mitch
Im pretty sure the codec produces samples and the samples are not converted to something that can have inter-sample peaks until D/A. So the DA needs a little headroom to handle this but it should not be an issue in the digital domain at all.
The codec can encode (sample) peaks above zero, and the decoders can handle these IF they are kept as floating point. However several of the biggest don't do this, they immediately revert to fixed point, with no headroom
PS. Now you mention it AAC is a slightly misleading choice for this demo since Apple's signal chain may actually be one of the few that uses floating point, but the issue is real
@@ProductionAdvice My understanding is that the audio card takes integer samples and during D/A the analog output signal can rise higher than where the integer samples are due to summing of the "sinc" functions used for D/A process (we cant catch every peak with a sample so some will for sure be between samples and they should be reproduced faithfully by the DA). A compressed signal, even a lossy one, should eventually be decompressed into integers (so should 32 bit floats be converted) before hitting the output device where the DA can produce intersample peaks. I think any software that is rendering the waves and resampling in a way that would produce the intersample peaks would need to handle that and have some headroom just like the DA would.. I suppose a buggy codec could do some bad math and truncate it, is that what you are saying is happening? That would be a serious issue and I'd be surprised if it were widespread.
@@williamtell1477 Right, I've been told by a very reliable source that some of the most popular streaming decoders render to fixed point immediately after decoding, before any other processing, including normalisation and gain. It's not buggy, it's by design (although I don't really understand why). We can hope it changes in future, but right now it's a thing.
@@ProductionAdvice I wonder if TruePeak metering during mastering avoids any such issues? That should ensure that everything ends up below 0dBFS even inter-sample peaks.
That hint hahaha thx for the video Ian!
to answer the title, yes it is a major problem
I just changed the title, sorry...! 😛
This was a really great video with a lot of great analysis. One thing I wonder though about the encoding clipping is whether it’s actually directly correlated to the intersample peaks? In other words, isn’t it possible that you can have a ISP of say -0.01dB and still get clipping after encoding losses?
Great question, and no, it's not that precise. But in my experience it's within half a dB or so. I master everything with -1 dBTP and it works pretty well. (I don't master super-loud, though)
@@ProductionAdvice Makes sense, thank you! Love your channel
Very informative, thank you, Ian!
My pleasure!
that was useful, thanks!
Pleasure !
not using true peak and oversampling in limiter is bold, i like how it saves transients and punch!
Thanks for this one Ian!
Pleasure!
One question, since the encoded file is stored in a floating point format, does that mean that if normalization is engaged, then those peaks might be preserved? Or is normalization occurring after the transition to fixed point?
It depends ! For example in my tests RUclips preserves peaks when accessed using Safari, but not Chrome...
I listened to on my phone on purpose and didn’t look at the screen for the listening test. I could hardly tell. Sometimes I was right but most of the time not.
That’s not surprising, this is way too subtle to be heard on a phone speaker. I think it’s audible on Apple AirPods though, for example
@@ProductionAdvice but that should be the worst case scenario and best place to hear it.
Yeah I guess AirPods are better but the phone is still a good testing medium
Great Video, do you have a link to the ABXer software that you showed?
It doesn't seem to be available right now, sorry :-/
Amazing! You stop saying dB´s. Correct, we do say 2 dB or two decibels. Nice video about ISP.
Very interesting. I frequently download popular songs from RUclips to study them and most of the time, the m4a file that I download has lots of peaks going above 0dB so when I normalize it to 0dB, the song goes down by 1 or 2 or even more dBs in some cases. Sometimes there are thousands of waves on each channel having this issue, going above 0dB. And these are top-charted songs possibly mastered by engineers among the best on this planet so would that mean that it's the big labels who still control how commercial music has to be mastered ?
Extremely loud masters are still the norm for most mainstream releases, unfortunately 😕
@@ProductionAdvice Thanks for fighting for the cause ! Step by step, let's make mastering great again !
Well YT is definitely adding to the volume for its YT music platform. (All TP measurements done in Izotope Rx 9). Bruno Mars 24K Magic 44.1kHz 16bit wav bought from iTunes shows -9.4 LUFS, +0.71 dBTP. M4A/MP4 HD Quality YT video download -8.5 LUFS, +1.85dBTP. YT Music streamed at -8.4 LUFS, +2.38dBTP. YT video streamed at -14 LUFS, -3.02dbTP. Stats for Nerds for 24K Magic shows -5.6 audio reduction so TP of the video should actually be -4.89dBTP, not -3.02. If I increased the original audio by 1 to match YT Music's inflated -8.4 LUFS, then the TP should read -3.89 (closer to -3.02). However, I'm a bit baffled why YT Music increased 24K Magic's stream volume by 1 LUFS, and its true peak by 1.67dB, when its highest quality codec 256kbps AAC should actually keep the volume the same and lower the TP from 0.71 to 0.54 (using Ozone 10's 256kbps AAC conversion). Even if I increased 24K magic's original audio by 1 dB, and then exported at 256kbps AAC, the LUFS becomes -8.4, but TP is still only +0.96dB. Maybe YT Music is doing some kind of expansion? And why do downloaded YT videos show the 1dB increase in LUFS as well?
@@Jrel Interesting analysis. I wonder what is the effect of this on what we hear when streaming music from RUclips. Does it affect quality negatively ?
Most of the songs I download from RUclips to study them are "boosted" above 0dB. First thing I do in the DAW is normalize and I have seen more than 2dB reduction happening in front of my eye in some cases while doing the normalize operation. This is a lot ! But when listening to the song on RUclips or even after I download it, to my ear it seems to sound good despite of this.
@@Jrel I know for a fact that RUclips doesn’t boost levels, either on the video or music platform. So the difference has to be in the uploaded files.
Since the difference is exactly 1 dB my guess is that this was mastered with True Peaks up to 0 dBTP, then reduced by 1 dB in an attempt to meet the Apple recommendation of -1 dBTP but without checking the actual decoded results. -8 is easily loud enough to make maintaining control of the peaks after decoding a challenge
The clipped ISP version sounds actually really "trashy" compared to the clean version.
Yeah, less clarity and space
Is there any particular true peak limiter you recommend?
I've had good results with the TC Electronic Brickwall and FabFilter Pro-L
@@ProductionAdvice fantastic. Thanks!
If you get 7/10 twice, you still only have 7/10. 14/20 = 7/10. I would say that it can be audible, probably isn't but since technically it is there.. it has to be part of the process to avoid ISP. I've used -1.6dB and see no reason to change, it is a game of probabilities after all and going past +1.6dB means someone has royally fucked up.. you don't get those unless you almost deliberately prepare a sample of "worst case scenario".
You're right that 18/25 is still only 72%, but it's much harder to achieve this by accident over a larger number of tests. So to be 95% confident I could really hear the differences (not a fluke) would require 100% success for 5 tests, or 90% for 10 tests, but only 70% for 25 tests. So based on this I'm now entirely confident that it's audible. (Well, 95 % 😛)
I agree that preparing for a worst-case scenario is the way to go, I master to -1 dBTP myself which has served me well.
I disagree about "royally f-ed up", though. This album was mastered by a top mastering studio, and there a huge number of other releases that have similar levels of ISPs after streaming. So this was a conscious choice - just not a great one, IMO !
Me recuerda a la voz de Fzst
Really in depth and thorough. Appreciate you and your amazing work! 🫡