[SoundAndMotion]

Here's another good video explaining bit depths and the effect on listening. The demonstration should be enough to convince the most ardent skeptics unless they just don't want to know.

The difference between 8bits and 16bits is noise, the signal is all there, nothing is missing. With dither, most people wouldn't be able to tell the difference between 8bits and 16bits for most commercial music, let alone 16bits compared with 24bits!

I like that video. Many similar videos come from a more engineering/signal theory perspective, but this video is something music lovers can appreciate. If one watches the whole thing and listens to what Mr. Shepherd says, a nice understanding of how bit-depth relates to music recording is available. Nice find.

One can't just pick out bits (lol) of what he says, though. Early on (0:42) he says "if you do it right", which he later (3:52) explains and demonstrates to mean adding appropriate dither. You do say: "With dither, most people ... most.. music," which is true. We agree. If you add noise shaping and oversampling, you can safely replace one of the "most"s with "all".

But the sentence "The difference between 8bits and 16bits is noise, the signal is all there, nothing is missing." is demonstrably false on its own. A passage hovering around -55 to -60dBFS will disappear if recorded at or truncated to 8 bits. So the signal is not all there. Am I being hypercritical or nit-picking? For most readers who'll never face this issue, yes. For the occasional DIYer who floats through here, and may have an Arduino board with 8-bit ADCs and some music project in mind, no, he/she must "do it right".

 

[castleofargh]

AFAIK, if quantization was done without dither, producing quantization error correlated to the input signal, this would be called a distortion. When quantization is done with dither, producing quantization error not correlated to the input signal, it is called noise.

The difference is most audible with quiet signals. Listen to the attached files:

• "02_8bit_no_dither.flac" sounds distorted when compared to the "01_16bit_orig.flac",
• "03_8bit_dither.flac" sounds the same, only with added noise, when compared to the "01_16bit_orig.flac",
• "04_8bit_dither_shaped.flac" sounds the same, with only a little noise, when compared to the "01_16bit_orig.flac".

(and why would you ever want to do quantization without dither is anyone's guess)

A word of warning about this. It's probably obvious for most who tried the files, but they are recording a very quiet signal that's close to and reaching into the 8bit cut. Which is why the hiss is really loud relatively to the music content even though it's 8bit down, and why the correlated noise(no dither) has so much impact and isn't masked much if at all by the music content.
Otherwise those are very nice examples.

 

[71 dB]

High bias cassette tape is equivalent to around 5bits digital and LP records around 10-11bits. The truth is even 8 bits is enough for most commercial music and for the bulk of listeners. After all, pre-recorded cassettes were the best selling music format from the late 1970s to the early 1990s.

I believe C-casettes can do much better than 30 dB of dynamic range when noise reduction is used. "Best selling" doesn't mean "hight quality." Effectively dithered 8 bit might be enough for some music such as commercial music with heavy dynamic compression, but we must use the most demanding music out there, not the easiest.

 

[old tech]

I like that video. Many similar videos come from a more engineering/signal theory perspective, but this video is something music lovers can appreciate. If one watches the whole thing and listens to what Mr. Shepherd says, a nice understanding of how bit-depth relates to music recording is available. Nice find.

One can't just pick out bits (lol) of what he says, though. Early on (0:42) he says "if you do it right", which he later (3:52) explains and demonstrates to mean adding appropriate dither. You do say: "With dither, most people ... most.. music," which is true. We agree. If you add noise shaping and oversampling, you can safely replace one of the "most"s with "all".

But the sentence "The difference between 8bits and 16bits is noise, the signal is all there, nothing is missing." is demonstrably false on its own. A passage hovering around -55 to -60dBFS will disappear if recorded at or truncated to 8 bits. So the signal is not all there. Am I being hypercritical or nit-picking? For most readers who'll never face this issue, yes. For the occasional DIYer who floats through here, and may have an Arduino board with 8-bit ADCs and some music project in mind, no, he/she must "do it right".

The signal is all there, even if drowned by noise. After all, DSD is 1 bit and the entire signal is recovered - albeit at very high sample rates and noise shaping. In any event, I think we are referring to playback rather than recording.

 

[71 dB]

But the sentence "The difference between 8bits and 16bits is noise, the signal is all there, nothing is missing." is demonstrably false on its own. A passage hovering around -55 to -60dBFS will disappear if recorded at or truncated to 8 bits. So the signal is not all there. Am I being hypercritical or nit-picking?

It's not false. The signal is all there, part of it is just masked by the noise. If you have the original 24 bit file from what the 8 bit version was made, you can produce the dither noise by substracting the 24 bit version from the 8 bit version (as done in the video). If you substact this noise (now in 24 bits) from the 8 bit version you get the original 24 bit version. All the signal is there. In music instruments mask each other to same extent, but that doesn't mean signal gets lost. It's all there, just partially masked.

 

[old tech]

I believe C-casettes can do much better than 30 dB of dynamic range when noise reduction is used. "Best selling" doesn't mean "hight quality." Effectively dithered 8 bit might be enough for some music such as commercial music with heavy dynamic compression, but we must use the most demanding music out there, not the easiest.

I don't disagree, just saying most listeners of the greater music consuming public would be satisfied. Most pre-recorded cassette tapes were not even high bias and the masses flocked to the dreadful early 128 MP3 codecs.

 

[71 dB]

The genius of dither is any 8 bit signal can pretend being 24 bit by pushing all the "last 16 bits stuff" onto the dither noise to do and when the dither can't do that because it's only 8 bits too, it gets truncated and badly distorted, but distorted noise is just noise. Nobody hears the difference. That's how we get rid of the distortion of loosing bit depth. Instead of distorting the signal we distort the noise were it doesn't matter. All we lose is dynamic range because the noise floor raises.

 

[71 dB]

I don't disagree, just saying most listeners of the greater music consuming public would be satisfied. Most pre-recorded cassette tapes were not even high bias and the masses flocked to the dreadful early 128 MP3 codecs.

You are correct. I just don't see the link to 24 bit vs 16 bit debate.
Those who think 16 bit is not enough are certainly not fans of pre-recorded cassettes or 128 kbps mp3s!

 

[SoundAndMotion]

The signal is all there, even if drowned by noise.

It's not false. The signal is all there, part of it is just masked by the noise.

[snip]
One can't just pick out bits (lol) of what he says, though. Early on (0:42) he says "if you do it right", which he later (3:52) explains and demonstrates to mean adding appropriate dither.
[snip]
A passage hovering around -55 to -60dBFS will disappear if recorded at or truncated to 8 bits. So the signal is not all there.

I tried to be clear that "doing it right" means adding dither. If you don't add dither to a signal that is -60dBFS, whether recorded at 8 bits or truncated to 8 bits from 16 or 24 bits, there is no signal. There's also no noise. At 8 bits, -60dBFS without dither is all zeroes!!
Where is the signal and where is the noise in all zeroes?

Edit: Also, I question the concept that a signal buried deep in noise is "there". There are tricks one can use, like signal averaging, dither and noise shaping, if one prepares ahead of time. But if I give you a file with a signal to which I've added enough noise to destroy the signal, the signal is no longer there in a real sense, perhaps in an abstract form (mathematically) or philosophically, but... really?
If a book is completely burned, and you save the ashes and the smoke, is the book still there?

 

[71 dB]

I tried to be clear that "doing it right" means adding dither. If you don't add dither to a signal that is -60dBFS, whether recorded at 8 bits or truncated to 8 bits from 16 or 24 bits, there is no signal. There's also no noise. At 8 bits, -60dBFS without dither is all zeroes!!
Where is the signal and where is the noise in all zeroes?

When your signal gets completely below the least significant bit, the quantization noise starts to correlate with your signal so much it's actually -100 % correlation. In other words your signal and quantization noise cancel each other completely and the result is silence.

 

[SoundAndMotion]

When your signal gets completely below the least significant bit, the quantization noise starts to correlate with your signal so much it's actually -100 % correlation. In other words your signal and quantization noise cancel each other completely and the result is silence.

That's philosophical drivel. The quantization error IS the signal. They don't both exist and cancel each other. The signal is not digitized at all, or is completely discarded (if truncated from a higher bit level).

 

[71 dB]

That's philosophical drivel.

So? It works and explains what happens. I find these philosophical drivels useful in understanding digital audio (and many other things).

 

[SoundAndMotion]

So? It works and explains what happens. I find these philosophical drivels useful in understanding digital audio (and many other things).

I have no problem with pedagogic tricks to help understanding something. You need to be careful with definitions of "noise" and "error", though. Go for it. But when you write (and others read):

The signal is all there, even if drowned by noise.

The signal is all there, part of it is just masked by the noise.

Arguing that a file of all zeroes is really the signal fully and completely there, except the noise has perfectly cancelled it to all zeroes, would defy anyone's definition of "noise".

 

[71 dB]

1. I have no problem with pedagogic tricks to help understanding something. You need to be careful with definitions of "noise" and "error", though. Go for it. But when you write (and others read):

2. Arguing that a file of all zeroes is really the signal fully and completely there, except the noise has perfectly cancelled it to all zeroes, would defy anyone's definition of "noise".

1. Sure, but in this context noise and error are the same thing.

2. What I said applies to the truncation process. 10 years later when everybody has forgotten why the data is all zeros it's just digital silence. The truncation process functions the same regardless of the amplitude of the signal to be truncated. What changes when the signal gets below the LSB of the truncated version is the error = noise in this context starts to correlate with he signal negatively in a very strong way. For example:

Signal = 0.25 LSB of 8 bit => Gets truncated to 0 LSB of 8 bits => error = noise = -0.25 LSB of 8 bit = -signal.

This is not only a "pedagogic trick." It's what happens mathematically. Error ( = noise) correlates with the signal and this correlation is signal amplitude dependent. The advantage of thinking this way is you don't need two models for signals strong enough to not truncate into zeros and signals so weak the result is all zeros.

 

[theaudiologist1]

I still prefer 48khz over 44.1 due to compatability. I think the 44.1 multiples should just be dicontinued.

Also, is there a point of releasing digital recordings recorded at 24bits to vinyl? Imo only analog recordings should be on vinyl. It's pointless.