Test Procedure for comparing 24 bit vs 16 bit audio

[Craig Maier]

I have not tried this yet, but as a result of discussions with Dan, I have written up this potential test to allow someone to compare the difference between 24 bit audio and 16 bit audio:


1. Start with a stereo 24 bits, 44.1 kHz file of good quality.

2. Convert it down to 16 bits using the change sample rate/resolution feature under the Edit Menu.

3. Up Convert that file back to 24 bits using the change sample rate/resolution feature.

4. Take that file and then apply the File Conversion Filter, with both channels checked for stereo to stereo and with inversion also check-boxed.

5. Paste Add this file to the original 24 bit source file.

6. What will be left over is just the difference signal between 16 bit and 24 bit audio.

Dan and Craig

[Dan McDonald]

Just did it.. the file is completely quiet and empty. I added 20dB to the sound (incresd gain) and still is quiet. Then I added 20dB more and there's something there... A whole lot of irritating noise.

[Craig Maier]

You are probably not surprised that I had concluded that from a different type of test that I had done sometime in this last year. I like to call these things (24 bit audio) the latest development in Marketing Technology. As they say one or two tests are worth 1000 expert opinions.

Craig

[Craig Maier]

The other test that I had run coming at the problem from a different angle is found here:

http://www.diamondcut.com/vforum/sho...highlight=bits

Craig

ps - looks like I ran the older test around February of 2014

[Dan McDonald]

Here's a picture of the difference, when I add 40dB of gain to the file. Just random noise.

[Craig Maier]

It's interesting and yet there will always be people that insist that they can hear the difference even if you perform some sort of scientific test. Many 24 bit advocates refuse to take such a test, interestingly.

Craig

[mikeebb]

Dan's results make sense. Bit depth fundamentally sets signal to noise ratio. 16 bit is (IIRC) theoretically about 96 db; 24 bit is well into the 100s. Of course, the sound source (if it's analog) will do worse than that - 60-70 db is doing pretty good for a used record or old non-master tape. So if he originally got silence (noise floor around -100 db) and added 40, he got it up to 60 db which is audible but not bad, again, for older analog sources.

I will use 24-bit when transferring good-quality analog to digital so I have some headroom for processing. I can keep the "recording" level down a bit to avoid clipping without worrying too much about the noise floor, and if noise builds up during processing it probably still won't get bad enough to mess things up. Then of course it goes to 16 bit for the CD burn. But in the real world there are so many other sources of noise (most amplifiers, even, probably can't do much better than 90-95 db s/n in a good but not super high-end system) that 24-bit just isn't necessary for the final product.

[Craig Maier]

To actually achieve 24 bits of resolution with a s/n greater than unity and in the audio band (20 to 20,000 Hz), one would need to cryogenically cool the frirst stage electronic devices to close to zero degrees K. We wrote this analysis up and placed it in the glossary section of the users guide.

Craig

[Marc Hildebrant]

It's useful to calculate the ENOB (Effective number of Bits) for your system when you are considering 24 bits or 16. The ENOB formula (find at Wikipedia) shows how the signal to noise ratio can degrade the "real" number of bits to a large amount.

I'm interested , though, in the comment about processing "overhead". I try to use the multifilter to minimize the any loss do to conversions "back and forth". Should I use 24 bits ?

Marc

[Craig Maier]

I just found it - it was under "Thermal Noise Floor" in the glossary section of the Diamond Cut helpfile. Here it is:

---------------------------------------------------------------------------------------------------------

Thermal Noise (Floor)

Any electrical conductor produces a random noise Voltage as long as it exists above 0 degrees K and/or has an electrical resistance greater than zero Ohms. The following formulae can be used to calculate the Root Mean Square value of the thermal noise Voltage of a terminating or source resistance:

E = (4RkT x ∆f) ^1/2 or E = √(4RkT x ∆f)

Wherein:

R = Resistive Component in Ohms (Ω)

k = Boltzmann’s Constant = 1.38 x 10^-23

Joules / Kelvin (1 Joule = 1 Watt x Second)

T = Absolute or Thermodynamic Temperature in degrees Kelvin

∆f = Bandwidth of the system in Hertz (Hz)

E = Root Mean Square (RMS) Noise Voltage

T (in degrees Kelvin) = Temperature in

Degrees C + 273.15

Example: Assume an audio mixer/microphone preamplifier is terminated with a 50K Ohm Resistance. It is operating at 40 degrees C internally, has a 60 dB Voltage Gain and exhibits a usable flat response from 20 Hz to 20 kHz. What is the RMS noise floor of the output of the mixer?

60 dB Voltage Gain = 1000 : 1 = 1000 X

Eout = (4RkT x ∆f) ^1/2 x 60 dB = ((4 x 50,000) x (1.38 x 10 –23) x (30 + 273.15) x (20,000 – 20)) ^1/2 x (1000)

Eout = 0.004 Volts RMS = 4.0 Millivolts RMS

It is important to note that this is the Noise Floor for this system and can’t be made to be any quieter than this number unless special techniques are employed. Cryogenic cooling techniques are sometimes used on first-stage amplification devices of specialized amplifiers to improve the noise performance of highly sensitive systems.

--------------------------------------------------------------------------------------------------------

Craig

[Craig Maier]

Use of the multifilter virtually eliminates rounding errors between filtering processes. But, for me anyway, I can only hear S/n values of around 14 bits, so it makes no difference to me in my restorations. Certainly, it does not hurt to use 24 bits.

Craig

[Craig Maier]

Marc made reference to ENOB (Effective Number of Bits) as it relates to S/n for A-D converters (or D-A converters). He cited a Wikipedia article on the topic. That article can be found at this link:

http://en.wikipedia.org/wiki/Effective_number_of_bits

Craig

[Marc Hildebrant]

Craig,

Thanks for posting the link. Note that the numbers are only correct if the noise is Gaussian.

Marc

[Craig Maier]

Marc,

Yes, I think that there was a denomonator that relates to that distribution (6.02). I guess that the coefficient could be changed if another distribution is known to exist on the noise.

Craig

[Dan McDonald]

Would this ENOB apply to requirements for an analog playback devices when converting to digital also? For example, if you have a cassette recording that you want to digitize and your player has a s/n ratio of around 65 dB or so, if you plug that into the formula, you get less than 11 bits needed. That would make sense to me but not sure if that's accurate.

Dan