On Sunday, 1 May 2022 at 16:31:41 UTC, claptrap wrote:

Autotune and vocal mixing are two different things, albide the normal population don't know the difference and think it's the same.

A lot of people mistake vocal mixing for autotune, when it really isn't.

Autotune takes vocals as input and changes each pitch to match a specific pitch etc.

Vocal mixing, might fix individual notes that were just sung wrong like an A that had accidenitally become A# a single time in the chorus and stuff like that, you don't go through all pitches in the vocal sample, on top of that it might add reverb, compression etc. all of which has nothing to do with autotune, but improves the sound a lot.

On Monday, 2 May 2022 at 01:43:03 UTC, claptrap wrote:

Technically even the Motorola 56000 can do over 500 FFTs per second with a window size of 1024 according to Wikipedia. So the phase vocoder part was feasible, but it might not have been sonically feasible in the sense that you would not end up with a product believed to be marketable or that it wasn't believed to be feasible to reach a sonic quality that would satisfy the market. That could come down to pitch-tracking, phase-vocoder issues or the details of putting it all together.

Phase vocoders do introduce artifacts in the sound, it kinda follows from the uncertainty principle, you get to choose between high resolution in time or high resolution in frequency, but not both. So when you modify the sound of chunks of sound only in the frequency domain (with no concern for time) and then glue those chunks back together you will get something that has changed not only in pitch (in the general case). So it takes a fair amount of cleverness and time consuming fiddling to "suppress" those "time domain artifacts" in such a way that we don't find it disturbing. (But as I said, by the late 90s, such artifacts was becoming the norm in commercial music. House music pushed the sound of popular music in a that direction throughout the 90s.)

However, the concept of decomposing sound into spectral components in order to modify or improve on the resulting sound has been an active field ever since ordinary computers were able to run FFT in reasonable time. So there is no reason to claim that someone suddenly woke up with this obvious idea that nobody had thought about before. It comes down to executing and hitting a wave (being adopted).

In general truly original innovators rarely succeed in producing a marketable product, market success usually happens by someone else with the right knowledge taking ideas that exists, refining them, making them less costly to produce, using good marketing at the right time (+ a stroke of luck, like being picked up by someone that gives it traction).

"Somone woke up with an obvious idea that nobody had thought about before" makes for good journalistic entertainment, but is usually not true. Successful products tend to come in the wake of "not quite there efforts". You very rarely find examples of the opposite. (The exception might be in chemistry where people stumble upon a substance with interesting properties.)

On Monday, 2 May 2022 at 08:52:06 UTC, Ola Fosheim Grøstad wrote:

The concept of "windowing" + "overlapp add" to reduce artifacts is quite old, e.g the Harris Window is 1978. Dont known for better ones (typically Hanning).
This doubles the amount of FFT required for a frame but you seem to say this was technically possible.

On Monday, 2 May 2022 at 06:15:32 UTC, FeepingCreature wrote:

Yay!

On Monday, 2 May 2022 at 08:52:06 UTC, Ola Fosheim Grøstad wrote:

Sometimes artifacts sound "good", be it for cultural or "objective" reason.

Many small delays can help a voice "fit in the mix", and spectral leakage in a phase vocoder do just that. So some may want to come through a STFT process just for the sound of leakage, that makes a voice sound "processed" (even without pitch change). Why? Because in a live performance, you would have those delays because of mic leakage.

It is also true of the artifacts that leads to reduced dynamics (such as phase misalignment in a phase vocoder). Didn't like those annoying vocal dynamics? Here is less of them, as a side-effect.

The phase-shift in oversampling? It can make drums sound more processed by delaying the basses, again. To the point people use oversampling for processors that only add minimal aliasing.

Plus in the 2020s, anything with the sound of a popular codec is going to sound "good" because it's the sound of streaming.

On Monday, 2 May 2022 at 00:24:24 UTC, Bruce Carneal wrote:

It does require new knowledge - you have to stick "macro" to the function declaration. In D, you don't need to do that, because the grammatical context the function is used in determines whether the function will be executed at compile time.

There is nothing big about CTFE. )

I need my beer badly right now!

On Monday, 2 May 2022 at 09:23:10 UTC, Guillaume Piolat wrote:

Yes, this is true. Like, the loudness-competition that lead to excessive use of compression (multiband?) and ducking (to let bass drum through) lead to a sound image that was pumping in and out. I personally find that annoying, but when you see kids driving in the streets playing loud music they seem to favour this "musically bad" sound. I guess they find excitement in it, where I think of it as poor mastering. And I guess in some genres it is now considered bad mastering if you don't use excessive compression.

I believe this loudness-competition and "overproduction" also has affected non-pop genres. If you get the ability to tweak, it is difficult to stop in time... I frequently find the live performances of talented singers on youtube more interesting than their studio recordings, actually.

The french music scene might be different? French "electro" seemed more refined/sophisticated in the sound than many other "similar" genres, but this is only my impression, which could be wrong.

I hadn't thought of that. Interesting perspective about mics, but a phase vocoder have other challenges related to changing the frequency content. How would you create a glissando from scratch just using inverse FFT, it is not so obvious? How do you tell the difference between a click and a "shhhhhhh" sound? The only difference is in the phase… so not so intuitive in the frequency domain, but very intuitive in the time domain. You don't only get spectral leakage from windowing, you also can get some phasing-artifacts when you manipulate the frequency content. And so on…

But, the audience today is very much accustomed to electronic soundscapes in mainstream music, so sounding "artificial" is not a negative. In the 80s you could see people argue seriously and with a fair amount of contempt that electronic music wasn't real music… That is a big difference!

Maybe similar things are happening in programming. Maybe very young programmers have a completely different view of what programming should be like? I don't know, but I've got a feeling that they would view C as a relic of the past. If we were teens, would we then focus on the GPU and forget about the CPU, or just patching together libraries in Javascript? Javascript is actually quite capable today, so…

I didn't understand this one, do you mean that musicians misunderstand what is causing the effect so that they think that it is caused by the main effect, but instead it caused by the internal delay of the unit? Or did you mean something else?

I hadn't though of that. I'm not sure if I hear the difference between the original or mp3 when playing other people's music (maybe the hi-hats). I do hear a difference when listening to my own mix (maybe because I've spent so many hours analysing it).

On Monday, 2 May 2022 at 08:57:21 UTC, user1234 wrote:

Yes, I assume anyone who knows about FFT also knows the theory for windowing? The theoretically "optimal" one for analysis is DPSS, although Kaiser is basicially the same, but I never use those.

I use 4x not 2x and Hann^2 (cos*cos) as the window function for simplicity. The reason for this is that when you heavily modify the frequency content you need to window it again. So you multiply with cos(t) twice, but when you add them together the sum = 1. Probably not optimal, but easy to deal with for experiments.

I also believe it is possible to use Hann-Poisson for analysis. It has excessive spectral leakage, but supposedly allows you to accurately find the peaks as the flanks for the spectral leakage are monotone (smooth slope) so you can use hill climbing. But I doubt you can use this for resynthesis.

What you could do is use Hann-Poisson for detecting peaks and then use another window function for resynthesis. I will try this some day :-).