 
|
Audio Asylum Thread Printer Get a view of an entire thread on one page |
For Sale Ads |
|
|
Audio Asylum Thread Printer Get a view of an entire thread on one page |
For Sale Ads |
76.90.39.24
In Reply to: RE: "44.1 actually sounds better than 24/192"... posted by Dave_K on October 08, 2015 at 03:21:47
I guess I don't understand "quantization".
It seems to me the higher the bit rate the less quantization is needed and less dithering to keep the distortion low so less noise?
Tre'
Have Fun and Enjoy the Music
"Still Working the Problem"
Follow Ups:
"It seems to me the higher the bit rate the less quantization is needed and less dithering to keep the distortion low so less noise?"The word length also reduces quantization.
What dither noise does is make the least-significant bit (LSB) switch randomly to signal levels or resolutions at amplitudes less than the LSB. When played back, with classical "sinc" filtering, the resultant level of the analog signal is close to the original signal below the LSB. Without the dither, the LSB would not be triggered randomly at all, the quantization error would then increase as a result.
With longer word lengths (ex: 20 bits instead of 16), with the extra bits, the levels "below the LSB" are now handled by those extra bits, the low level signal is tracked more accurately than without the extra bits and using dither.
Edits: 10/08/15
It seems to me the higher the bit rate the less quantization is needed and less dithering to keep the distortion low so less noise?
Right.
With truncation, the magnitude of the quantization error is < 1 LSB, and with rounding, it's < LSB/2. And to linearize the quantization process, the required dither is at the level of the LSB. The smaller the LSB, the lower the dither noise. More bits = lower digital noise floor.
The truth is I don't understand what the quantization process is.
How can the amplitude of each sample be made correct without an infinite number of bits?
How does the system know how much to "fudge" (and in which direction) the amplitude of each sample?
Thanks.
Tre'
Have Fun and Enjoy the Music
"Still Working the Problem"
Let's see if I have this right.To me the quantization process is not a process at all, it's the outcome of a less than perfect system.
When you digitize an analog signal, while using a word length that is less than infinite, quantization takes place and creates distortion.
Dithering will lower the harmonic distortion orders down into an increasing noise floor than has been deemed to be acceptable.
So now back to what I was trying to say.
In absolute terms, the higher the bit rate the less, distortion causing, amplitude inaccuracies there will be.
It begs the question, how high of a bit rate would it take to listen to the stream without dithering while having acceptable HD numbers?
Tre'
Have Fun and Enjoy the Music
"Still Working the Problem"
Edits: 10/08/15 10/08/15
Let me give you the long answer.The analog signal that you are sampling contains the musical signal S(t) plus an analog error signal consisting of distortion D(t) plus noise N(t). Digital sampling without dither turns this into sequence of quantized sample values x(i) which encode the analog signal plus error value (S+D+N) at time t(i) along with the addition of a second error signal Q(i). The quantization error Q is the difference between the value of S+D+N and the nearest digitally representable signal level. So x(i) = S(t(i)) + D(t(i)) + N(t(i)) + Q(i).
Because Q is a residual of rounding the sum of S+D+N to the nearest digital value, it is correlated to the sum of S+D+N. The spectrum of Q will depend on the spectra of those components and on how well it correlates with each of those S,D,N components, which in turn depends on the level of those components relative to Q. For example, if S> D> N> Q or S> N> Q> D, then Q will be mainly correlated with N and will have a noise-like spectrum. Since it's lower in level than N, it's buried in the analog noise floor and you can consider the sampling process to be effectively transparent. On the other hand, if S> Q> D> N or S> Q> N> D, then Q will be mainly correlated to the signal S. In that case, it will be a type of distortion whose spectrum is a function of the signal and the sampling rate. Since it is also above the analog noise floor N, the sampling process cannot be considered transparent.
For 16-bit sampling, Q is at the -96 dB level, which may be above or below the analog noise floor depending on the nature of the input signal (tape, mic feed, MIDI, etc.) So there are cases involving low noise input sources where 16-bit sampling without dithering can produce a spectrum of distortion products that many people agree is audible and everyone agrees is undesirable.
Dithering in analog to digital conversion is the intentional mixing of a pseudo-random analog signal into the input of the sampler at a level slightly above Q. So x(i) = S(t(i)) + D(t(i)) + N(t(i)) + dither(t) + Q(i). A typical dither signal used in audio is at the level of +/- 1 LSB and has a triangle PDF. It has a flat white-noise like spectrum. It's level is 3 dB above Q, so Q will end up being correlated with the dither rather than the signal, and therefore Q will have a white-noise like spectrum. The tradeoff is slightly more noise but no added distortion.
A second benefit of dithering is that periodic signals which are below the quantization noise floor are preserved in the sampled data and can extracted by time averaging. Without dithering, in the cases where Q> N, those signals are turned into distortion by the quantization process and lost.
For 24-bit sampling, Q is at the -144 dB level, which is pretty much guaranteed to be below the analog noise floor of any current recording equipment. So dithering is optional in a 24-bit ADC but it is usually done anyway. However, if you take 24-bit data and reduce it to 16-bits to fit on a CD, you still need to dither (in a digital form) to avoid introducing distortion. More generally, any process that results in a reduction in bit depth should also include dithering. Many operations performed in the digital domain by a mixing or mastering engineer require multiplication of digital values, a process which results in a longer word length than the input (e.g. multiplying two 24-bit numbers can product a result that is up to 48 bits in length). These results have to dithered rather than simply truncated or rounded to fit back into 24 bits. So most digital audio operations involve dithering of intermediate results.
To finally answer your question: With a minimalist production chain, you could conceivably record, produce, and release at 24-bits with no dither used at all. But the more processing you perform, the greater the likelihood of some quantization distortion reaching an audible level. Which is one of the reasons why dithering is a de facto standard in modern digital audio regardless of the end user delivery format.
Edits: 10/09/15
This answer and your other answer plus some more reading on my part and I now have a different understanding of the whole subject.
I am now left with the question, why doesn't digital sound better than it does?
I used to work at a recording studio. We had a Ampex ATR 102. We also had a 24/96 digital system using Apogee I/O's (circa 2005).
Comparing the live mic feed to the ATR (properly aligned using 456 at +3 and 30ips no Dolby) was almost indistinguishable.
The digital never did that. At least not to my ears.
If the blame doesn't belong where I was placing it....where does it belong?
Whatever, I'm going to go play an old, made from a analog master tape, vinyl record. :-)
BTW I just read the link below, circa 1998. 24/96 digital to analog converter, 8x oversampling filter.
Tre'
Have Fun and Enjoy the Music
"Still Working the Problem"
I am now left with the question, why doesn't digital sound better than it does?
...
If the blame doesn't belong where I was placing it....where does it belong?
An issue with the CD format in particular is that the sample rate is too low.
And a general issue with digital audio is timing, i.e. jitter.
Beyond that, it's probably integrated circuit design. The sonic performance of audio op-amps is all over the map, despite having similar specifications. Same for ADCs and DACs.
Some will blame RFI. I'm less convinced of that.
.
Have Fun and Enjoy the Music
"Still Working the Problem"
+1 ! Thanks, Dave K.
That was an excellent summary/description of the main important points/concepts of digitizing a signal.
Much of what you wrote and explained is stuff which I knew, but you pulled it all together into one tidy little package, like few people can do. Great job on your post!
Higher sample rates yield a more "focussed" digital image of the analog signal, and higher bit depths yield a more accurate representation of "soft to loud". Quantization error makes perfect sense to me. But dither has been a question which you've helped to explain.
It's not surprising that pro studios use high bit rates and high bit depths when recording tracks - there's often a lot of processing afterward.
:)
:)
Post a Followup:
| FAQ |
Post a Message! |
Forgot Password? |
|
||||||||||||||
|
||||||||||||||
This post is made possible by the generous support of people like you and our sponsors:
Top of Page ]
[ Contact Us ]
[ Support/Wish List ]
[ Copyright © 2024, Audio Asylum, all rights reserved ]
