24bit vs 16bit, the myth exploded! | Page 157 | Headphone Reviews and Discussion - Head-Fi.org

Greenears · Feb 3, 2015 at 11:09 PM

rrod said:
I'd be careful testing that one. What was the definition of instantaneous DR at a sample?

Both hands firmly over ears ....

stv014 · Feb 4, 2015 at 6:47 AM

greenears said:
Thank you so much for the file - I downloaded it fine. Unfortunately on this network right now I (a) don't have a Unix prompt and (b) don't have gcc installed. I guess I could download it for Windows, or borrow a friend's system to try. Not a big barrier I'll try it at some point in the next couple of weeks.

It can be built with basically anything that can compile a simple "hello.c" program. A basic installation of MinGW should work. I could also create and upload an .exe version.

I assume the 44 byte .wav format above is the standard one that comes out of SOX.

44 bytes is the minimum length of the WAV header. This is a very simple program with no library dependencies, so it does not actually parse the header of the input file, it just copies the first 44 bytes unchanged, and assumes that the sample format is stereo 16-bit PCM. For converting a 24-bit input file, it will need to be modified.

castleofargh · Feb 4, 2015 at 7:25 AM

how do I go about to try that bit extermination when I'm a noob and don't have linux? I did something like that with adobe audition(or was it cool edit?) what seems like an eternity ago, but my years thinking I should crack any software just because I could are gone(like my hair). so if possible I would like to do it legit this time ^_^.
or have somebody risking jail for me by uploading a few different bitdepth values of some nice dynamic classical track?

(is it legit if I'm not the one going to jail?)

Quote:

greenears said:
rrod said:

I get about 60dB or so for the the LOTR track (taking the 99th percentile of 2s RMS values and the max peak). I don't know how LKFS loudness works (I assume it's meant to work in concert with standards for delivery systems in theaters), but it seems like it means more like "you can't have a louder perceived sound than -31dB full scale" rather than a dynamic range measurement.

Click to expand...

I think everyone needs to look at this post and be reminded of something key: All DR measurements require a window of time and a method of averaging (in this case time window is 2s and the method is 99 percentile etc) The instantaneous DR (1 sample time) can always be much higher as has been pointed out about 20 pages ago. I'm told that most modern tracks are normalized to peak around -6 or -3 dBFS you basically could have 90 dB instantaneous DR with redbook.

The real question I and others are asking is (a) how much of these instantaneous peaks are there (b) are there enough to become audible (c) can you tell the difference if you extend them to 120dB DR (20 bits equivalent). Another reference pointed out that a short peak of 120 dB is not noticeable, even though longer exposure (seconds?) will cause pain.

I think the video posted by limpidglitch makes the overall DR value vs fullscale peak, clear to people who are not familiar with all kinds of DBs and measurements.
the -3db or more I believe (but pro dudes may have more ideas about that), is a simple safety against clipping in general. I know that I've had some mp3s that seemed to clip, and lowering the gain (whatever the way) removed that feeling. maybe one day I'll remove the gain values of everything just to find out what tracks seemed to do that an analyze the crap out of them.
anyway if I was making a track, I wouldn't want people to think it's crap simply because they have a DAC with too much voltage output, or some upper rounding math trick clipping stuff close to zero db when oversampling, or just a guy who doesn't know how to use an EQ. better safe than sorry.

stv014 · Feb 4, 2015 at 11:18 AM

This updated version of the quantize program includes a Win32 executable, and supports 24-bit PCM samples (files created by sox and the dsputils programs should work). When processing 24-bit input, the output file is always in the same format, so it does waste some space. Another change is that adding 4 to the dither type switches from rounding towards the nearest integer to floor rounding.

RRod · Feb 4, 2015 at 11:31 AM

stv014 said:
This updated version of the quantize program includes a Win32 executable, and supports 24-bit PCM samples (files created by sox and the dsputils programs should work). When processing 24-bit input, the output file is always in the same format, so it does waste some space. Another change is that adding 4 to the dither type switches from rounding towards the nearest integer to floor rounding.

\m/

unshavenbastard · Feb 4, 2015 at 3:51 PM

And 1GHz is a much higher frequency than 20kHz, but so what? Your ears can't hear it and your ears also can't hear sounds at -144dB, which are represented by the 24th bit in digital audio. Your ears can't even hear the 18th bit. This is all wishful thinking, and proven as such in controlled tests.

Inaccuracies like these is what irritates me about this thread.
The original poster made the same mistake, claiming all there was to 24 vs. 16 bits was dynamical range.
Well from what it looks to me anyway. I'll try to explain, maybe I myself misunderstood something.

People here seem to confuse dB with dB-audio, the latter being an absolute measure (because it has an absolute reference, not depending on device), the former not.
dB just says how much louder it is compared to a reference level, the dynamic range of e.g. a soundcard being how much louder the loudest value is from the most quiet one.
But this could be scaled / mapped onto a differently spaced grid(e.g. when you fumble with the volume dial of your amp), not changing the dynamic range being used, as not only the loudest amplitude of a recording will be louder, but also the most quiet one (so the absolute audio level range changes, of course).
E.g.: if your DAC has a nominal dynamic range of 8bits or 48dB let's say outputting 1mV..255mV, and you externally amplify so that you get values from 25.5mV..6553.5mV, it's still 48dB, but your speakers put it into a different dbA range, and onto a different place within the human hearing range.

Nobody mentioned the one thing that would be deciding whether more bits can offer more perceived quality or not: The resolution of the ear itself. And it is indeed about resolution, not just range.
You could make a DAC which can produce only two different output levels, and space those such that they span 100 dB, so you have a "dynamic range of 100dB", but only two steps, which is obviously useless. It's about how many different amplitude values within the healthy dynamic range of the ear the latter can discern, similar to color formats' bit depths and the number of different shades of colors the eye can discern.
So more finely grained steps along the amplitude axis can matter, as long as it does not exceed the number of amplitude steps the ear can discern within the region of the overall hearing range of sound pressures that this granularity is mapped onto.
If, hypothetically, the ear could discern 1 bilion different audio levels within the 140 or so dB*Audio* level range of the ear, you could easily map those 144 dB range of a 24bit DAC onto, say, the center third of the ear's dbA range, and the ear could still discern all the steps - i.e. if there were less steps resolution there, the ear might notice the equivalent of visual banding in low number of color shades images!

Note that I am not making any specific claims about bit rates and their supposed advantages here.
Just generally speaking, the explanations given in this thread so far seem inaccurate.

From what I remember (can't find a reference) the ear actually has a resolution of about 17 bits, which would mean it could discern twice as many different audio levels than 16bit wave can. So 24 bits would be 128 times overkill - assuming that during playback it is all amplified perfectly into just the sound pressure range of hearing where it matters.
How this all really works out when people playback their music, I'm not sure ^^
.

castleofargh · Feb 4, 2015 at 4:27 PM

unshavenbastard said:
And 1GHz is a much higher frequency than 20kHz, but so what? Your ears can't hear it and your ears also can't hear sounds at -144dB, which are represented by the 24th bit in digital audio. Your ears can't even hear the 18th bit. This is all wishful thinking, and proven as such in controlled tests.

Click to expand...

Inaccuracies like these is what irritates me about this thread.
The original poster made the same mistake, claiming all there was to 24 vs. 16 bits was dynamical range.
Well from what it looks to me anyway. I'll try to explain, maybe I myself misunderstood something.

People here seem to confuse dB with dB-audio, the latter being an absolute measure (because it has an absolute reference, not depending on device), the former not.
dB just says how much louder it is compared to a reference level, the dynamic range of e.g. a soundcard being how much louder the loudest value is from the most quiet one.
But this could be scaled / mapped onto a differently spaced grid(e.g. when you fumble with the volume dial of your amp), not changing the dynamic range being used, as not only the loudest amplitude of a recording will be louder, but also the most quiet one (so the absolute audio level range changes, of course).
E.g.: if your DAC has a nominal dynamic range of 8bits or 48dB let's say outputting 1mV..255mV, and you externally amplify so that you get values from 25.5mV..6553.5mV, it's still 48dB, but your speakers put it into a different dbA range, and onto a different place within the human hearing range.

Nobody mentioned the one thing that would be deciding whether more bits can offer more perceived quality or not: The resolution of the ear itself. And it is indeed about resolution, not just range.
You could make a DAC which can produce only two different output levels, and space those such that they span 100 dB, so you have a "dynamic range of 100dB", but only two steps, which is obviously useless. It's about how many different amplitude values within the healthy dynamic range of the ear the latter can discern, similar to color formats' bit depths and the number of different shades of colors the eye can discern.
So more finely grained steps along the amplitude axis can matter, as long as it does not exceed the number of amplitude steps the ear can discern within the region of the overall hearing range of sound pressures that this granularity is mapped onto.
If, hypothetically, the ear could discern 1 bilion different audio levels within the 140 or so dB*Audio* level range of the ear, you could easily map those 144 dB range of a 24bit DAC onto, say, the center third of the ear's dbA range, and the ear could still discern all the steps - i.e. if there were less steps resolution there, the ear might notice the equivalent of visual banding in low number of color shades images!

Note that I am not making any specific claims about bit rates and their supposed advantages here.
Just generally speaking, the explanations given in this thread so far seem inaccurate.

From what I remember (can't find a reference) the ear actually has a resolution of about 17 bits, which would mean it could discern twice as many different audio levels than 16bit wave can. So 24 bits would be 128 times overkill - assuming that during playback it is all amplified perfectly into just the sound pressure range of hearing where it matters.
How this all really works out when people playback their music, I'm not sure ^^
.

no you have the wrong idea about resolution or increased precision. well you're right if we only look at values for being values, so on the digital side, you're right.
but wrong when you look at it as soundwaves. when you end up with the DAC making the sine wave to output the analog signal, it will be a continuous signal so it won't matter that you have fine tuned each sample or not with a precision up to 17 or 24bit. the maximum difference between the wave generated from the 24bit signal and the 16bit signal will be some noise 16bit down and below, as that's the LSB in 16bit and de facto, the biggest error in precision you should expect(at least in theory).

meaning that of course you're increasing the precision of the analog signal by using 24bit, but the difference is audible only below -96db. and it works like that in part because of how waves can mix together. the wrong signal can be seen as the wrong signal, or as the right signal+some added noise. looking at it that way, you see that you end up with the right signal from 0 to -96db, and then you have a mess as loud as the errors from the DAC(or signal's bitdepth) so as loud as 16bit down = -96db noise. and that's quantization noise. we're back on our feet.

RRod · Feb 4, 2015 at 4:38 PM

unshavenbastard said:
[text]
.

As far as dynamic range, it is intimately related to the number of bits, as a lower number of bits simply cannot represent the same minimal and maximal waveforms that a higher number of bits can (techniques like noise-shaping notwithstanding). Take 16 vs. 8 bits and square waves. The maximal amplitude symmetric (around 0) square wave you can get at 16bits has its peaks at ±32767, while the minimal has peaks at ±1. That's a 90dB difference in amplitude. At 8bits, you get maximal peaks at {0,254} and minimal at {126,128}, which is a 42dB difference in amplitude. So we can get 48dB more difference between peak values by going to 16bit for this waveform.

Resolution, as you are using it, would come down to being able to detect the smallest change possible for a given signal. So for square waves that would mean, at 16bit, detecting the difference between ±32767 and ±32766. Maybe I'll make that example.

castleofargh · Feb 4, 2015 at 5:02 PM

rrod said:
unshavenbastard said:

[text]
.

Click to expand...

As far as dynamic range, it is intimately related to the number of bits, as a lower number of bits simply cannot represent the same minimal and maximal waveforms that a higher number of bits can (techniques like noise-shaping notwithstanding). Take 16 vs. 8 bits and square waves. The maximal amplitude symmetric (around 0) square wave you can get at 16bits has its peaks at ±32767, while the minimal has peaks at ±1. That's a 90dB difference in amplitude. At 8bits, you get maximal peaks at {0,254} and minimal at {126,128}, which is a 42dB difference in amplitude. So we can get 48dB more difference between peak values by going to 16bit for this waveform.

Resolution, as you are using it, would come down to being able to detect the smallest change possible for a given signal. So for square waves that would mean, at 16bit, detecting the difference between ±32767 and ±32766. Maybe I'll make that example.

and I was afraid my explanation wouldn't be clear enough for him ^_^.

RRod · Feb 4, 2015 at 5:06 PM

castleofargh said:
and I was afraid my explanation wouldn't be clear enough for him ^_^.

I'm not all good with the baking analogies as you are :frowning2:

Here are two files at 16/44100. One is a square wave with amplitude 255; the other with amplitude 254 (software can easily distinguish them; they are about 0.03dB RMS apart).
https://drive.google.com/file/d/0BwmVtb5IwniESGV5ODdvZ3pROUk/view?usp=sharing
https://drive.google.com/file/d/0BwmVtb5IwniES3RNakVoLV9CUmc/view?usp=sharing

lamode · Feb 4, 2015 at 7:00 PM

unshavenbastard said:
Inaccuracies like these is what irritates me about this thread.
The original poster made the same mistake, claiming all there was to 24 vs. 16 bits was dynamical range.
Well from what it looks to me anyway. I'll try to explain, maybe I myself misunderstood something.

People here seem to confuse dB with dB-audio, the latter being an absolute measure (because it has an absolute reference, not depending on device), the former not.
dB just says how much louder it is compared to a reference level, the dynamic range of e.g. a soundcard being how much louder the loudest value is from the most quiet one.
But this could be scaled / mapped onto a differently spaced grid(e.g. when you fumble with the volume dial of your amp), not changing the dynamic range being used, as not only the loudest amplitude of a recording will be louder, but also the most quiet one (so the absolute audio level range changes, of course).
E.g.: if your DAC has a nominal dynamic range of 8bits or 48dB let's say outputting 1mV..255mV, and you externally amplify so that you get values from 25.5mV..6553.5mV, it's still 48dB, but your speakers put it into a different dbA range, and onto a different place within the human hearing range.

Nobody mentioned the one thing that would be deciding whether more bits can offer more perceived quality or not: The resolution of the ear itself. And it is indeed about resolution, not just range.
You could make a DAC which can produce only two different output levels, and space those such that they span 100 dB, so you have a "dynamic range of 100dB", but only two steps, which is obviously useless. It's about how many different amplitude values within the healthy dynamic range of the ear the latter can discern, similar to color formats' bit depths and the number of different shades of colors the eye can discern.
So more finely grained steps along the amplitude axis can matter, as long as it does not exceed the number of amplitude steps the ear can discern within the region of the overall hearing range of sound pressures that this granularity is mapped onto.
If, hypothetically, the ear could discern 1 bilion different audio levels within the 140 or so dB*Audio* level range of the ear, you could easily map those 144 dB range of a 24bit DAC onto, say, the center third of the ear's dbA range, and the ear could still discern all the steps - i.e. if there were less steps resolution there, the ear might notice the equivalent of visual banding in low number of color shades images!

Note that I am not making any specific claims about bit rates and their supposed advantages here.
Just generally speaking, the explanations given in this thread so far seem inaccurate.

From what I remember (can't find a reference) the ear actually has a resolution of about 17 bits, which would mean it could discern twice as many different audio levels than 16bit wave can. So 24 bits would be 128 times overkill - assuming that during playback it is all amplified perfectly into just the sound pressure range of hearing where it matters.
How this all really works out when people playback their music, I'm not sure ^^
.

You are theoretically correct about the distinction between resolution and DR. I used to try and explain this to photographers who thought that a 16-bit/channel image had higher DR than the same file converted to 8-bit. No, the DR was the same but the increments were less smooth.

But... all of this becomes moot as no proper double blind test has ever revealed a single test subject who could tell the difference between a live audio feed, and the same feed fed through a 16/44 AD-DA loop, which is evidence enough for me that although many problems exist in theory, in real life they are inaudible.

Greenears · Feb 4, 2015 at 9:50 PM

stv014 said:
This updated version of the quantize program includes a Win32 executable, and supports 24-bit PCM samples (files created by sox and the dsputils programs should work). When processing 24-bit input, the output file is always in the same format, so it does waste some space. Another change is that adding 4 to the dither type switches from rounding towards the nearest integer to floor rounding.

Cool - thanks I downloaded it. I don't have time now but I'll try it out later.

Greenears · Feb 4, 2015 at 10:08 PM

lamode said:
You are theoretically correct about the distinction between resolution and DR. I used to try and explain this to photographers who thought that a 16-bit/channel image had higher DR than the same file converted to 8-bit. No, the DR was the same but the increments were less smooth.

But... all of this becomes moot as no proper double blind test has ever revealed a single test subject who could tell the difference between a live audio feed, and the same feed fed through a 16/44 AD-DA loop, which is evidence enough for me that although many problems exist in theory, in real life they are inaudible.

All this back and forth on DR can be settled by this new utility posted. Simply keep chopping of a few bits at a time and till you pass the ABX test, then run the track through a DR analyzer.

I think people will find the standard equation of DR = N bits * 6 in dB only works at the limit where DR is at the max of N*6, but not at lower values. You'll be able to hear noise at levels lower than the equation predicts.

But .... I've been wrong before :--) Or have I?

castleofargh · Feb 4, 2015 at 10:49 PM

greenears said:
lamode said:

You are theoretically correct about the distinction between resolution and DR. I used to try and explain this to photographers who thought that a 16-bit/channel image had higher DR than the same file converted to 8-bit. No, the DR was the same but the increments were less smooth.

But... all of this becomes moot as no proper double blind test has ever revealed a single test subject who could tell the difference between a live audio feed, and the same feed fed through a 16/44 AD-DA loop, which is evidence enough for me that although many problems exist in theory, in real life they are inaudible.

Click to expand...

All this back and forth on DR can be settled by this new utility posted. Simply keep chopping of a few bits at a time and till you pass the ABX test, then run the track through a DR analyzer.

I think people will find the standard equation of DR = N bits * 6 in dB only works at the limit where DR is at the max of N*6, but not at lower values. You'll be able to hear noise at levels lower than the equation predicts.

But .... I've been wrong before :--) Or have I?

you just gave me the audiophile idea of 2015. making a box that you plug between the DAC and the amp that will add tape hiss. you will be able to set the value so that it's slightly above both the amp noise floor and the DAC bit depth. the end of THD, IMD etc, pure hiss over all frequencies just a tad louder to cover it. analog dithering!!!!!!!!!!
I guess 3500$ the box is a good starting price? maybe I should really put a tape in it so people would have to turn it from time to time like a true audiophile. and then we launch a market for different tapes with different hisses. hiss rolling is the new tube!!!!!!!!!
man I'm on fire. and the worst part is that with an ok marketing that's the kind of stuff that could probably sell and make the sound more "natural".

stv014 · Feb 5, 2015 at 7:24 AM

Originally Posted by unshavenbastard /img/forum/go_quote.gif

You could make a DAC which can produce only two different output levels, and space those such that they span 100 dB, so you have a "dynamic range of 100dB", but only two steps, which is obviously useless. It's about how many different amplitude values within the healthy dynamic range of the ear the latter can discern, similar to color formats' bit depths and the number of different shades of colors the eye can discern.

Actually, it is possible to implement high quality audio playback with only two output levels. That is how DSD works, and most modern DACs also have only a few bits of output resolution when playing PCM. The use of dithering (which can be psycho-acoustically optimized with noise shaping, so that most of the quantization noise is at frequencies where the threshold of hearing is high) allows for encoding any fractional level at the cost of adding noise to the signal.

The dynamic range of an audio device is the ratio of the highest output level it is capable of without clipping and the noise floor. Therefore, if the dithering is well implemented, the equation of 'X bits = X * 6.02 + Y dB of dynamic range' (Y is a constant depending on the dithering/noise shaping used, and possibly other factors like the use of A-weighting, and the sample rate) is indeed generally correct for high enough values of X. It can be proven that by adding triangular distribution white noise with a peak to peak level of 2 LSB before the quantization, as long as clipping is avoided, any input level will result in the same output level on average, with the addition of white noise at a constant 0.5 LSB RMS level.

Originally Posted by unshavenbastard /img/forum/go_quote.gif

From what I remember (can't find a reference) the ear actually has a resolution of about 17 bits.

It was probably calculated from an assumed "good enough" dynamic range. Other than that, the concept of "bits of resolution" in the time domain cannot really be applied to hearing, as audio is not perceived in PCM format. However, the smallest audible change in the amplitude of a tone is actually about 1%, or 0.1 dB. Lossy compression algorithms successfully take advantage of this limited resolution by quantizing (after normalization) the signal in the frequency domain, even to a very low number of bits when masking makes it possible without audible artifacts.

Latest Thread Images

Greenears

100+ Head-Fier

stv014

Headphoneus Supremus

castleofargh

Sound Science Forum Moderator

stv014

Headphoneus Supremus

RRod

Headphoneus Supremus

unshavenbastard

New Head-Fier

castleofargh

Sound Science Forum Moderator

RRod

Headphoneus Supremus

castleofargh

Sound Science Forum Moderator

RRod

Headphoneus Supremus

lamode

500+ Head-Fier

Greenears

100+ Head-Fier

Greenears

100+ Head-Fier

castleofargh

Sound Science Forum Moderator

stv014

Headphoneus Supremus

Users who are viewing this thread