Jump to content
IGNORED

Is recording at high bit depth objectively better?


Rexp

Recommended Posts

On 2/21/2021 at 10:53 PM, erin said:

The objective reply is that 16 bit recordings have 96dB dynamic range.

 

With dithering, they can achieve 100dB dynamic range.

 

Subjective comment: surely this is enough dynamic range to capture 99% of music recordings and definitely enough dynamic range to capture recordings from vinyl records.

 

Subjective comment:

In my experience, there is nothing wrong with 16 bit recordings. I have never heard any dynamic compression due to 16 bits being used.  Any differences in sound between 16 bit and 24 bit recordings are due to the way the DAC chip manages the audio conversion from D to A. Any sonic difference is caused by the hardware, not the actual 16 bit data.

 

The real value for higher bit depth or higher sample rates is about signal processing, not really just listening.

 

When you start with 16 bits and work only in 16bits, any gain less than 0dB will start affecting the dither.  Also, pure 16bit processing can create zipper noise with certain situations.   Likewise, higher sample rates are useful for dynamics processing -- that is, any fast gain control creates sidebands - it is best to  do processing at double the normal sample rate (4X the maximum signal frequency) because of modulation effects.   You can often get by with 3X the maximum signal frequency (I know, off topic, but trying explain the real reasons for higher rate/bit depth.)

 

Of course, with pure 16bit processing, there are dynamic range issues other than truncating the dither in extreme cases also.

 

This is why, when reading in a 16bit file, if there is any processing, then upconvert to 24bits/32bits or floating point first.   Convert back down when done.   Same goes for sample rate -- if you are doing dynamics processing, then do an immediate sample rate upconversion from 44.1k (or even 48k) to at least 66k for audio (88.2k/96k are better.)

 

Doing signal processing purely at the delivery rate/bit depth (44.1k/16 bits), without due consideration, can really cause troubles.   Delivering at the 44.1k/16bit  rate is probably okay though.  (I prefer 48k or higher even for delivery, but dont really have a strong argument for it.)

 

 

 

 

 

Link to comment
1 hour ago, Rexp said:

I find the majority of commercial Redbook stuff unlistenable and was thinking its either the ADC used by some studios or the DSP/mastering at fault. Guess it could be either but recording at higher bit rates doesn't seem essential. 

My new decoder (with my new headphones -- the old ones were giving me too much bass, therefore telling me to make the output thin.)


The decoder WILL demonstrate the cause of bad sound.   (I am again hearing what I had been happy with, but with working headphones.)

I don't know when the old headphones failed -- but must have been 6mos ago.

 

Link to comment
  • 2 weeks later...
1 hour ago, bluesman said:

The ADCs I own are all 24 bit fixed point.  The input is converted to 32 bit floating in the computer.  As I understand it, 24 bit signals map into 32 bit float well.  A 32 bit float with normalization and size bit gives the equivalent of 25 bits of mantissa.  This doesn’t change the signal going into the computer - it improves the process of recording, editing, processing etc and yields a better stereo output.
 

I still have to be careful with signal levels from individual mics and instruments.  Where it helps is in the master bus.  As I add more individual tracks, the master level goes up and I have to bring it down successively as I lay down more tracks.  With 32 bit float recording, minor peaks above 0 dB in the master track do not require reducing the gain.  I used to have to stop recording and cut the gain because the master would clip if I didn’t check and correct the input volume before recording the next track. 

 

I can post-process the 32 bit tracks with impunity and use the stereo master directly for export as a 16 or 24 bit wave or FLAC.  This is more useful for live recording than for vinyl ripping, but I set my DAWs to default to 32 bit. So I use the same settings for ripping.

Yes --

certain kinds of processing and mixing like to use more bits than just 16.   Certain kinds of processing like rates at least 1.5X to 2X normal Nyquist rate (at least 66.15k if done carefully, or 88.2k/96k if some leeway is needed.)  I suggest just 96k and be done with it.   Sometimes even some processing might benefit from 192k, but I haven't run into my self.

 

FP is 'the way to go' though for processing -- eliminates a lot of internal SW troubles.

 

Link to comment
On 3/8/2021 at 8:47 AM, davide256 said:

The thing to remember about CD rate is that it was a compromise based on assumptions around dynamic range  and audible distortion from sampling rate. Most of the

time it works well if engineered correctly for dynamic range. Lately I've begun to suspect that 48/24 is all thats really needed and that higher rates are just to

aid the less than perfect electronics we all struggle with.

I agree with you 100% about 48k/24 for distribution and listening.   Certain processing can require higher rates, but I don't really think that we are focusing on production issues.  (Lossless and trivial rate conversion can be done if processing is needed -- so 48k/24 is even a technically plausible rate for professional interchange because integral rate conversion is essentially perfect and totally trivial to do.)

 

Frankly, 44.1k makes me nervous, but in the margins, works well enough.  But, why not 48k except for CDs?   And about the #bits, why not 24bits nowadays?  Even if the benefit really stops at 17-18 good bits, 24bits is nice, even number.  Of course, like the sample rate fraud that goes on, it is easy to create false, simply dithered bits to sell 'higher quality'.

 

So I agree, 48k/24, is the best tradeoff in my opinion.

 

 

Link to comment
10 hours ago, Rexp said:

Maybe Dither is what I object to? Mastering engineers use DAW's that operate at 32bit float. Any down conversion will cause truncation unless Dither is applied. 

Dither can be problematical during processing.  It is easy to make shaped (or any) dither disappear by filtering or EQ, so that is one reason why 16bit processingis best avoided.   Myself, I cannot hear proper dither for 16bits, but perhaps some people can.

 

You REALLY need dither when converting to 16bits, but it is NOT a panacea, esp during processing.

Think about this -- you have a 16bit system, then do -6dB gain control -- where did most of the dither disappear to?   Gotta upconvert to more bits to do ANY processing, or do your processing in analog after conversion from digital.

 

John

 

Link to comment
11 minutes ago, Rexp said:

My conclusion is 16bit is fine for recording but since all audio editing is done at 32bits+ it would be better to distribute music at 32/44.1

Yes -- 16bits is okay to record, but remember that any gain control really needs to be done at higher res to keep the dither intact.   Once all gain control is done before committing the data (e.g. setting levels), the converting down to 16bits is perfectly okay.  Dither MUST happen before converting to 16bits from higher res.  Dither is super critical and must come from higher res than you are committing, or it doesn't help.  Dither basically solves the 'stair stepping' problem that DOES happen in some ways unless dither isn't used.   The signal STILL has troubles if dither is attempted directly on a 16bit signal, unless the target res might be 14bits.

 

Distributing music would be good enough for perfection at 44.1k or 48k/24 -- simply because the levels can be controlled for produced music, and +1 and -1 limits aren't a problem.   There is literally no trouble at 44.1k/48k 32bit float  also, and gives the flexibility to handle peaks >1.0 and differences between  low level signals.

 

Where 32bit float is really good -- no dither required.   Also, it can handle out of range values.  Since normal 16, 24, 32bit integer (even though you don't find 32 bit integer very often, except inside of SoX) that they are limited to +-1, and at super low levels, there is loss of precision, then the FP standard is probably best for ALL production and *even* secreted *INSIDE* all consumer equipment.   A limited 4bit exponent/20bit mantissa would be okay inside of consumer equipment when done purely in hardware instead of on a CPU.  The reason for the 20bit precision (20/4bit) instead of 16/8 is that dither is better maintained with a 20bit precision scheme.  The dither *can* be maintained with a 16/8, but a 20/4 gives enough extra dynamic range while making dither trivial to maintain.   ANYWAY -- in most pro applications, 32bit FP 24/8, is good enough and maintains dither really well.

 

My decoder sometimes needs 64 bit float to maintain extreme precision at extreme varying levels.  This is because the brute force math that I do has wide dynamic range, and I am subtracting two very close numbers to get a result that must be low distortion.  My case is odd though, and even though some DAW software uses 64bits, it normally isn't really needed.   There are sometimes more intelligent ways to do what I am doing, perhaps getting around the need for 64bits, but my project is already so complicated -- I need to chip away at the use of 64bits, when possible, by improving algorithms, then convert to 32bits.

 

John

 

 

Link to comment

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now



×
×
  • Create New...