I don't know too much about dither but I was reading about it. So I'd like to discuss some issues that I was pondering.

First, is the difference a good dither algorithm makes audible? This thread apparently has people claiming so. And looking at the reaction in this thread I thought it is apparently not. Also I remember the discussion on whether checking the option "Dither" in the MPC plug-in makes an audible difference and the responses were almost like "If you don't have extreme hearing abilities, no".

Now also about the dithering process: As far as I could understand the dithering is a problem associated to quantization (Compare these:
24 bit 100Hz Sine Wave, 24 Bit 100Hz Sine Wave Truncated to 16 Bits, 24 bit Sine Wave with Dither Noise Added (RPDF), 24 bit Sine Wave with Dither Noise Added (RPDF) Truncated to 16 bits). Thus as soon as we quantize the harmonics are affected therefore there's a noticeable difference in the spectragraph. From this point on we intend to get as close to the original signal as possible. But whenever dither is mentioned we consider our signal to be still in the digital domain. Comparing different dither algorithms we observe the spectragraph plot of the dithered 16 bit wave to see how successful we are.

However at this point shouldn't we be actually observing our success by looking at the spectragraph after the Digital to Analog conversion? Because after all the harmonics are going to change again. Because the jumps at consecutive sampling points (like this wave) would be somewhat smoothened out. The reason is we can't have almost discontinuous jumps in waves in nature (I am not sure how correct I am here). Therefore our D/A converter and our equipment is going to produce a smoother transition at the jumps and our wave that we observe is going to be much different than what we observe on the spectrograph. And that's what we hear and what we want to improve on. Therefore I was wondering why no one observes the reproduction quality of the original wave after the D/A conversion (Compare the spectrograph plots of Wave 1 and Wave 2 in : Wave 1-> A / D 24 bit -> Dither algorithm -> Truncation to 16 bit -> A / D conversion -> Wave 2)

Also wouldn't the dither and digital to analog conversion process be more successful in reproducing the original wave if we increase the sampling rate (naively by introducing new sampling point in between for instance that averages the previous consecutive samples) and then
applying the dither?

I was wondering whether any hardware device that is built to or could dither after the digital to analog conversion step. I asked this question before here, but couldn't understand Pio2001's response exactly.

I don't have much time to answer, so I'll be concise.

As to different dither types being audible, take into account that in that thread you point, the dither noise was amplified several tenths of dB in order to make it clearly audible, in practice its audibility would be dubious. Look at my last post at there.

Dither is someting you do in digital domain, before going to analog. Dither in analog has no sense. Dither allows to remove quantization distortion that happens anytime there is some requantization going on, or, what is the same, anytime there is a conversion to a lower resolution format (from 24 bit to 16 bit, or from high- res float to fixed point, for example). Using proper dither there won't be any quantization distortion in this process, not using it there can be (depending on the signals being converted). So, it is advisable to use it always.

If you don't use dither and get quantization distortion, this distortion will be there after going to analog, and there will be no way to remove it once it's there. Hence, dither has to be done before going to analog.

I meant "Couldn't we achieve the same effect (which is to reduce the effect of quantization) by some other analog device instead of dithering??" in that thread.

But my other question is, when comparing the efficiency of dither algorithms, shouldn't we compare the frequency analysis/spectrograph of signals after the D/A conversion? Maybe some dither algorithm would perform better and would get a closer approximation to our original signal (I refer to "Wave 1" in my example) after the D/A process thanks to other physical phenomena related to D/A conversion and audio reproduction equipment? Therefore examining the spectragraph plot in the digital domain does not make perfect sense to me...

F*ck, I just wrote a nice reply explaining all that but did a wrong "click" and lost it. You'll have to wait until I feel like writing all that again, sorry.

Until that, take a look at those documents:

http://www.benchmarkmedia.com/appnotes-d/ditherap.pdf

http://www.mediarte.org/download/software/...io/whitepapers/ , last doc.


You can take a look also here:

http://www.hydrogenaudio.org/forums/index....&f=1&t=5884&hl=


Sorry for the bad language used.

To notice a difference (talking about 16bit resolution) you need
- music recorded at low volume (at least for some seconds)
- decent equipment (-> SNR)
- extreme hearing abilities will help too


Smoothing out would mean lowpassing the signal which should not take place. To get an impression of what's done in D/A conversion have a look at Waveform view of a decent Sound Editor. You'll see that the waveform goes through every single sample. Between the samples the curve is smoothed/rounded - by this it is ensured that the resulting waveform only contains frequencies below Nyquist limit (22.05kHz if sampling rate is 44.1kHz).


Think of clicks due to wrong (=discontinuous) samples on playback of badly scratched CDs or transients in music ... Besides, dither noise has an max. amplitude of 1 (flat) or maybe 20 (strong noiseshaping), while the signal's amplitude is usually more than 1000 times bigger (= + 60dB). So you can hardly call that "discontinuous jumps".


Again, I don't think so as smoothing would mean lowpassing in this case. Even if this would be true, you can still use spectral view of a dithered digital signal. Just disregard the frequency range above your asumed lowpass.

Of course. The higher the samplig frequency the more headroom you have to apply inaudible noiseshaped dither and lower noise floor in the audible range. In this case applying a lowpass (starting at ~20 kHz) while/after D/A conversion is a good idea. IIRC SACD can be considered as an extreme example for this (2.8 MHz sampling rate / 1 bit resolution; more information in [url=http://www.hydrogenaudio.org/forums/index.php?act=ST&f=20&t=3390&hl=sacd]this thread[/b].)


In analog domain there are no discrete steps/values, so quantization errors can't occur -> dither is useless (just adding noise). If quantization distortion has been introduced in digital domain there's no possibility to pick the added frequencies and get rid of them in analog domain.

The goal is not to get the closest possible approximation to the original signal (waveform) - If you add the differences ("original" - 16 bit digital) off all sample values, you'll see that truncation gives a lower total error than dithering.
The goal is to avoid distortion (= added certain frequencies caused by digital steps that occur in a regular pattern) and to preserve audible information below LSB level.


All "physical phenomena related to D/A conversion" that do something else than perfect reproduction of the waveform as it can be seen e.g. in Cool Edit Waveform View (Zoom in ...) must be considered as flaws. If the flaws of a D/A converter are known, it's probably possible to process the signal before D/A conversion to "balance" these flaws (e.g. by resampling to 48kHz using a decent equipment before handing the signal over to a crappy soundcard or by using an equalizer), but this has nothing to do with dither.