February 16, 2023 at 1:23 pm #4929
If you have created a dynamic, well-shaped, balanced master, that will be submitted to a digital aggregator and converted to lossy formats, and it has a LUFS level of -14 or less, and a TP that is at or less than the required -1.0, do you still need to use a limiter to be safe even if it’s not doing any limiting?
February 16, 2023 at 10:25 pm #4933James JohnstonParticipant
You’re also evaluating for intersample overs? If so, you should be safe, except in the most unusual situations. To be almost 100% absolutely safe you need your highest peak (including intersample) to be approximately at -4dBFS. Such situations have to be very nearly contrived, however.
February 16, 2023 at 11:27 pm #4934
Hi James. Yes, I check for intersample overs. When you say “To be almost 100% absolutely safe you need your highest peak (including intersample) to be approximately at -4dBFS” how does that jive with the peak of -1.0 dBFS recommended by digital aggregators? Which, I understand, prevents distortion from occurring when converting from 24/44.1 or 16/44.1 to the compressed formats that are used by streaming services.
February 17, 2023 at 6:54 am #4937
The very few situations JJ mentioned about the highest peak do have to be “carefully contrived” as JJ mentioned. For example, I recorded my talk/demonstration for the SF section of the AES. And at the end asked for a figurative round of applause for one of the attendees. And I applauded three inches in front of a Shoeps hypercardioid condensor mike. Later, while sweetening the pcm recording for later re-streaming I examined the level of the 45 minute recording (made without a limiter and including music demonstrations and speech) and found that somewhere in it there was a +11 dBTP! What what what? Where where where?
Sequoia’s algorithm for finding the highest sample peak in the recording totally failed to find the spot. I figured there must be some failure in the true peak algorithm. I mean really “+ 11 dBTP”!!! I divided the search through the recording in half and in half again and again until I isolated the highest sample peak to be —— aha!!!!! That pesky applause.
The applause was so transient and its peak level so instantaneous that inserting my trusty DMG Limitless did nothing practically. It was too fast for the true peak detection limiter to even react or show. Pushing the threshold radically did get it to react of course in the most hilariously sick sounding way you can imagine. And clearly this section of the recording was anything but essential. Using Sequoia’s object-based processing I cut the object with the crazy applause and applied the digital limiter only in that section just for shits and giggles and maybe some concerns that it might cause a lossy codec to turn over and die 🙂
But seriously, this is what JJ meant by a contrived situation. In most cases a 4X oversampling true peak measurement or detection algorithm will find what you want and you can either manually attenuate if necessary to arrive at -1 dBTP max, which is usually satisfactory for any reasonably speedy lossy bitrate. Or peak limit if it’s a good sounding invisible limiter. Or let it clip the codec if it’s a mild clip and not a sonically very important part of the work, like this close miked applause.
So I concur with you working without a limiter for 99% of the music you’ll encounter in the wild. If the TP doesn’t exceed -1 dBTP for 99% of that typical material.
I can confirm that -1 dBTP is a good number to not exceed for You Tube’s Opus 128 Kbps codec with high quality music. I performed a test with some “audiophile quality” music and it really went down hill if I let it get to 0 dBTP max. And -1 was acceptable.
And considering Spotify is a little more forgiving and Apple’s lossy now can do 48 kHz at 256 Kbps VBR I would say you are probably fine for most music without a limiter up to -1 dBTP.
February 17, 2023 at 2:33 pm #4946
Thanks Bob! I was smiling and chuckling through your whole story. And thanks for putting my mind at ease about not using a limiter. When I mentioned it to a couple of mix engineers who do rock and metal, they looked shocked. “You don’t use a limiter? Are you nuts? How do you get tracks to be loud enough then?” I said, ummm, I just turn up the gain on my stereo system until my ears are bleeding and the dog leaves the room. Then I know it’s loud enough.
February 17, 2023 at 10:43 pm #4953
I dunno why applause would kick up the true peak 11 dB, but a tricky problem we get in EE grad school is this (assume our unit of time is the same as the sampling period, so by definition, the sample rate is 1).
So what is shown is perfect reconstruction of continuous-time x(t) from the samples x[n] where it looks like it was a pure tone at the Nyquist frequency. But at t=0, there is a polarity reversal. It turns out that, even though all of the samples are limited in amplitude to ±1, the reconstructed x(t) blows up to infinity between t=-1 and t=0.
So this applause would have to be some pretty goofy set of samples toggling polarity each sample instance with one hiccup where they don’t toggle polarity.
February 17, 2023 at 11:29 pm #4954
This is where it gets over my head quickly, but I believe the extreme true peak reading has something to do with the upsampling that’s used to extrapolate the true peak level.
February 18, 2023 at 12:23 pm #4958
Well, now I gotta write a MATLAB program to show this. It’s kinda cool. Give it a day.
UPDATE: It’s gonna take me a little longer. My MATLAB chops are rustier than I first thought. Representing signals having different sample rates in MATLAB is a pain in arse. But my plan is a nice pic showing how the true peak of an amplitude-limited digital signal can shoot up a few dB using nearly ideal interpolation.
March 4, 2023 at 8:55 pm #5088Alexey LukinParticipant
Using the standard 4× upsampling filter from BS.1770, one can’t get more than +6.2 dB of true peak overshoot, even on RBJ’s signal. However the standard does not prescribe the single correct way to measure true peaks, so different meters may apply various amounts of oversampling. For example, RX would show +7.4 dBTP on the same signal, as it exceeds the minimal 4× upsampling requirement.
March 5, 2023 at 9:25 pm #5093
Hi Alexey! I’m really glad to see you here.
The nasty signal I was illustrating is one where none of the samples exceed the ±1.0 rails, yet the “perfectly” reconstructed analog signal, x(t), has an overshoot that goes to infinity at t=-0.5 because the infinite sum of 1/n is unbounded and the polarity is never toggled in the sum (it’s all positive 1/n). If your upsampling filter is long enough, your +6.2 dB overshoot will be exceeded.
March 5, 2023 at 9:36 pm #5095Alexey LukinParticipant
Hi Robert! I fully agree: an infinite filter will give you an infinite overshoot on such a signal.
March 6, 2023 at 12:05 am #5101
Here I finally wrote and ran a dumb MATLAB script. (I actually need folks like Alexey to motivate me to finish projects that I start.)
Here’s the frequency response (dB vs. log frequency) of the brick-friggin-wall filter:
March 6, 2023 at 12:06 am #5102
And here are the original samples (little red circles) and the nearly “perfect” reconstructed analog signal (blue):
All of the samples lie exactly on the rails, but the interpolated signal goes past it, more than 14 dB over the rails. This is a ridiculously long windowed-sinc function for the brick-friggin-wall impulse response. 4096 samples. So it was looking at 2048 samples into the past and 2048 samples into the future to interpolate between any two adjacent samples. Note the interpolated curve always goes through the original samples.
March 7, 2023 at 10:04 am #5112
Bravo, Alexey and RBJ! Way over my head but I understand the gist: In layman’s terms, it’s possible to construct an artificial signal whose sample peak does not exceed 0 dBFS but whose true peak can be much much higher. Yet still I have no idea how Sequoia’s true peak algorithm found the applause to reach +11 dBTP. I haven’t lost any sleep over it, however 🙂
Another hint: The more distorted, bright and compressed the music you master, the greater the chance it will produce 0 dBFS+ signals and totally mess up the sound, especially once it gets to a codec.
March 7, 2023 at 1:14 pm #5114
Bob, thanks for the translation! I think I get it, sorta kinda maybe…LOL. And a big thank you to the math geniuses Alexey and RBJ who took the time to illuminate us mere mortals and expose the magic behind the tools we take for granted.
You must be logged in to reply to this topic.