What is lost in computer recording compared to original tape?

[xh3rv]

Just finished watching the video. He does well on explaining the stair step myth. Now can we address my question? His video does not.

My question is

How can you gain something that was never recorded though? If the original wav has gaps? A analog signal does not have the same gaps. Even though we can't hear them without slowing it down?

There really are no gaps in the digital version, not just when sampled/reproduced but also in theory. Samples come with discrete values of time but they represent a continuous segment of time (and in a particular way: http://en.wikipedia.org/wiki/Dirac_delta_function). With a high enough sampling rate, sampling is both gapless and frequent enough to capture the relevant features of an analog waveform.

[e] The video with one of the Ableton time-stretch algorithms is really off. The glitch isn't gaps between samples, but a consequence of granular deconstruction. One could get the same glitching with a robot scratching vinyl.

[Josmoker]

Nice video, unfortunately the guy is a bit off in some passages.
- the timestretch-example is completely wrong - if you slow down a record the sound is played back at a slower speed and at the same time the pitch goes down. This is perfectly possible in digital. What he describes is timestretch, which is not possible in analog - but also in this case the dots are used more than once, if you will (hence the "stuttering"), rather than pulled apart.
- "more dots" aka a higher sampling frequency does not mean higher quality but a higher maximum sampled frequency. Basically 2 dots per cycle are enough, so with regard to the graph in his video, he has more than enough dots to accurately represent his wave already.

[memyselfandus]

Cool. How do you slow down digital audio a lot.. without hearing gaps? And artifacts? With say a human voice with no effects added. Been trying to figure out how to do this. From my understanding it's not possible because of the snapshots. What is the source of the artifacts we hear when doing extreme time stretching in digital?

[lfm]

There really are no gaps in the digital version, not just when sampled/reproduced but also in theory. Samples come with discrete values of time but they represent a continuous segment of time (and in a particular way: http://en.wikipedia.org/wiki/Dirac_delta_function). With a high enough sampling rate, sampling is both gapless and frequent enough to capture the relevant features of an analog waveform.

Putting it really simple might go something like:

The filter taking away the sampling frequency itself - will smooth out any stairs in there. Stairs will be more like smooth hills.

A stair is basically a squarewave part - and contain in theory a series of odd harmonics of the sampling frequency. And all that is going away(or in practical terms reduced) enough by the crossover filter(or what it might be called in english).

Every octave matters - so running 96k instead of 48k means another 6,12 or 24dB removal of sampling frequency - or instead using less steap filter allowing for less phasing artifacts in high end.

Dithering
------------
You got this at least 24-bit recordings done, and now you are to reduce to 16-bit CD quality.

24-bit allow for +/- 8,000,000 steps.
16 bit allow for +/- 32 000 steps.

So each stair is much higher in 16-bit.

Having signal going from a certain level to another will change one stair up or down.
This happening in the same place a lot in signal will soon form a pattern - and ears create signals from patterns.
And since stairs are fewer - it will easier form an audible pattern.
So to make signal not jump a stair on the same spot every time - you introduce some noise(certain shape) on a low level(a part of a stair height).
This will make it more like random transition, assuming noise is random, and what would have been an audible digital artifact - goes away.

[Josmoker]

It depends what you are after. You could do it the same way as you do it with analog and accept the downpitching. Or you accept it sounding stretched - a good algorithm shoud be able to avoid an audible stuttering, but the audio will not sound natural anymore at a certain point.
The only way to get around this would be resynthesis - to extract the properties of the voice speaking and what is spoken, and reproduce it at the required rather then the original speed, rather than slowing down the audio.
Then again musicians have always been inspired by the limitations of any (at the time being) current technology.

[VariKusBrainZ]

You can time stretch in analog domain using bucket brigade delay lines

[Ogopogo]

How can you gain something that was never recorded though? If the original wav has gaps? A analog signal does not have the same gaps. Even though we can't hear them without slowing it down?

You do lose something in the gaps, but since audio is fundamentally different than video it works different. All you lose are frequencies so high that they can't fit in the gaps, in other words, squiggles in the audio signal that would fit between the gaps, and for modern digital systems, that means frequencies above the human hearing range.

When the digital to analog conversion occurs, it's sort of like a connect the dots process, where the the digital samples are the dots. So just like you can recreate a drawing by doing this connect the dots process, the da convertor recreates the wave. You might say that a connect the dots drawing never really looks right because the lines aren't perfect, but I don't believe this is an issue for da conversion. Not sure exactly how that works but due to the nature of what it's recreating it can recreate it exactly, at least it recreates it exactly up to the point where the frequency is too high to fall within that dots.



So the problem with this graph is that there really isn't anything such as "digital wave" that you can compare to an analog wave. The digital info is just info thats needed to recreate analog waves, just the dots themselves. Your daw shows it as a wave but that's just for editing purposes. And you'll notice that when you really zoom in, different daws represent the waveform in different ways since there isn't exactly a right way to do it. Some will show you the dots themselves not being connected, some will show the dots and they'll show a line running through them that represents how the wave will be after analog conversion, some will show you the stair steps. But all that's really there is the dots.

[kvaca]

You can time stretch in analog domain using bucket brigade delay lines

maybe in theory...or you do know any device capable of this??

[xh3rv]

You can time stretch in analog domain using bucket brigade delay lines

maybe in theory...or you do know any device capable of this??

http://www.wendycarlos.com/other/Eltro-1967/index.html

This info from Howard Scarr in the Uhbik manual

[e] Well, it's a tape-based device but, time-stretch with no pitch-shift, which in a sense requires pitch-shifting time-stretched audio ... dunno if there's a BBD device that does this. If the broader question is, are at least analog devices that time-stretch ...

[memyselfandus]

Tons of good info guys. Thanks. Basically digital leaves shit out and adds some to the story what it thinks is missing? Sorta like a bad ex wife? All 1's and zero's?
Seriously.. Thank you for taking the time for all the replies guys.

I need a nap. Looks like Reaper 5 is coming.

[memyselfandus]

You can time stretch in analog domain using bucket brigade delay lines

maybe in theory...or you do know any device capable of this??

http://www.wendycarlos.com/other/Eltro-1967/index.html

This info from Howard Scarr in the Uhbik manual

[e] Well, it's a tape-based device but, time-stretch with no pitch-shift, which in a sense requires pitch-shifting time-stretched audio ... dunno if there's a BBD device that does this. If the broader question is, are at least analog devices that time-stretch ...

Now we're talking

[kvaca]

You can time stretch in analog domain using bucket brigade delay lines

maybe in theory...or you do know any device capable of this??

http://www.wendycarlos.com/other/Eltro-1967/index.html

This info from Howard Scarr in the Uhbik manual

[e] Well, it's a tape-based device but, time-stretch with no pitch-shift, which in a sense requires pitch-shifting time-stretched audio ... dunno if there's a BBD device that does this. If the broader question is, are at least analog devices that time-stretch ...

it is not BBD based, but still very interesting /if somehow limited/ design!
would be interesting to know how many semitones up or down /when pitchshifting/ or how many % /when time-stretching/ it was capable compared to digital?

[aciddose]

A BBD is not analog unless you consider a DAC with lots of jitter to also be analog in time. The BBD has no amplitude quantization, but its very purpose is time quantization. It is a series of sample&hold circuits, which are also used as the input to a working DAC.

Only instead of going in to the DAC (which performs amplitude quantization / conversion to binary / delta-sigma) it is a chain of many S&H circuits one after the other.

The important thing to remember is that the signal only moves forward on the clock edge (depends upon the chip) and so at that point a "sample" is taken in time and thus time is quantized, meaning it is no different from a fully quantized sample other than that the amplitude is continuous.

As I mentioned already, a S&H is the part that does the "sampling" for an ordinary DAC also, there is no difference.

[aciddose]

Regarding the discussion about how to display samples when "zoomed in", actually there is a correct way: sinc.

That is sinc(x) = sin(x) / x

Each sample is reconstructed fully when it is replaced by sinc.

Everything we can actually do merely approximates it, but as far as simply drawing lines on the screen even a four-point spline can visually approximate sinc very well at common resolutions. If you were to construct a tool to compare the two (which I may do now actually, since I already have it, I'd just need to build/write a sinc interpolation template) you might find you need to zoom in a lot to notice any significant difference.

The smallest difference you could notice would be 256 (8-bit color) times the number of vertical pixels used. Let's assume 1080, 1080 * 256 = 276480. So if I'm doing my math correctly, we need sinc(276480)...

1 / 276480 = -108.8db

So we get better than we need for 16-bit accuracy anyway at 1080 vertical resolution using an insane window width like 262144 (half, so 524288 full window), which is slightly less than 276480 but still gives -108.37db.

Anyway I have my doubts any of us can see a difference in intensity at the edge of a line drawn on the screen anywhere close to 1/256th, 1/64th would be more reasonable which would give 1080*64 = 69120. Close enough to 65536 to round there which is -96.3db, probably more on-target considering the quantization noise in a 16-bit signal and our visual acuity.

What we could actually be bothered by? I doubt sinc is very useful compared to a four-point spline.

[memyselfandus]

Why are there artifacts when you slow down digital audio? even just a voice with no effects? what is that?