testing one two... 16bit 44.1kHz vs 24bit 96kHz

[JohnRoberts]

An old friend is taking a survey about bit/sample rate.

Comparing 16bit 44.1kHz and 24bit 96kHz files

http://ethanwiner.com/hd-audio.htm 

enjoy

JR

 

[abbey road d enfer]

Often discussed subject. Advocates of the "no-difference" expose the same old argument: "digital audio is continuous without steps. The only thing bit depth affects is the level of background noise."
This is simply not true. Not very wrong either, but it has to be put into context, which is the mathematical concept of continuous signal, for which this affirmation is true. There is a formula that qualifies the amplitude error of a sample as a function of the Signal frequency to sampling frequency (as the ratio increases, the error increases too, for 100% error when signal frequency=SR).
Put differently, for less than 1% error, the SR must be >22times the signal frequency. But this formula is for a single occurence of the signal; error decreases significantly with the number of samples. Not too dissimilar with the constraints of FFT.
Let's not forget that audio is not a continuous signal.
The subject is made a little more complex by the fact that most converters are Delta/Sigma, which again involves dealing with strings of data that are interrelated.
Many people have conducted their experiments on the subject; most reasonable reviewers (I exclude audiophools) come to the same conclusion. There is no perceptible difference between 16 and 24 bits as a release format, except when dealing with poorly recorded material or when extensive processing is required. Single Speed (44.4 and 48k SR) is inadequate, mostly because of the constraints on the anti-alias and recovery filters; Double Speed is more than adequate (in fact, if they had a time machine, many designers would chose 60-64k). Quad Speed has no operational advantage, though it is of some use to those who do measurements.

 

[JohnRoberts]

Often discussed subject. Advocates of the "no-difference" expose the same old argument: "digital audio is continuous without steps. The only thing bit depth affects is the level of background noise."
This is simply not true. Not very wrong either, but it has to be put into context, which is the mathematical concept of continuous signal, for which this affirmation is true. There is a formula that qualifies the amplitude error of a sample as a function of the Signal frequency to sampling frequency (as the ratio increases, the error increases too, for 100% error when signal frequency=SR).
Put differently, for less than 1% error, the SR must be >22times the signal frequency. But this formula is for a single occurence of the signal; error decreases significantly with the number of samples. Not too dissimilar with the constraints of FFT.
Let's not forget that audio is not a continuous signal.
The subject is made a little more complex by the fact that most converters are Delta/Sigma, which again involves dealing with strings of data that are interrelated.
Many people have conducted their experiments on the subject; most reasonable reviewers (I exclude audiophools) come to the same conclusion. There is no perceptible difference between 16 and 24 bits as a release format, except when dealing with poorly recorded material or when extensive processing is required. Single Speed (44.4 and 48k SR) is inadequate, mostly because of the constraints on the anti-alias and recovery filters; Double Speed is more than adequate (in fact, if they had a time machine, many designers would chose 60-64k). Quad Speed has no operational advantage, though it is of some use to those who do measurements.

Are you debating yourself?

Ethan just published some blinded audio files and asked people to vote which they thought were the higher resolution formats. I don't think he wants to prejudice the results.

I've known Ethan since the 70s. He built and operated his own recording studio (including building his own recording console) and is a musician with above average ears himself.

JR

 

[abbey road d enfer]

Are you debating yourself?

I don't think so. The sentence I quoted is from the link. I'm just pointing at a worn-out argument that, as presented, tends to indicate that the author has a certain bias.

I've known Ethan since the 70s. He built and operated his own recording studio (including building his own recording console) and is a musician with above average ears himself.

I know who Ethan Winer is; that's why I'm a bit surprized by his affirmation, that is an oversimplification of a complex subject.

 

[Matador]

Let's not forget that audio is not a continuous signal.

Care to elaborate on this?  Most formal definitions for continuous-time signals, like:

A continuous-time signal is an infinite and uncountable set of numbers, as are the possible values each number can have. That is, between a start and end time, there are infinite possible values for time t and instantaneous amplitude, x(t).

and:

A signal of continuous amplitude and time is known as a continuous-time signal or an analog signal. This (a signal) will have some value at every instant of time. The electrical signals derived in proportion with the physical quantities such as temperature, pressure, sound etc. are generally continuous signals. Other examples of continuous signals are sine wave, cosine wave, triangular wave etc.

Audio (at least as a representation of sound pressure traveling through air) would seem to fit these definitions, at least between some sensible range of input values.

 

[JohnRoberts]

I don't think so. The sentence I quoted is from the link. I'm just pointing at a worn-out argument that, as presented, tends to indicate that the author has a certain bias.

sorry I do not see a quote broken out separately. After re-reading Ethan's linked page I do not see the offensive comment. 

OK after re-reading it several times you object to his comment about word length and noise( that I now see you italicized).  FWIW Ethan is not a digital theory expert, but more of a practical recording expert, with a strong BS filter.

I know who Ethan Winer is; that's why I'm a bit surprized by his affirmation, that is an oversimplification of a complex subject.

I didn't read his comments that closely, I think he was just setting up the rationale behind publishing the listening test samples. 

It will be interesting to see what results he gets. He has already deleted some posts from his facebook page where people shared their results, so as not to not bias others.

I expect him to report the results fairly and he tried to engineer out any tells so people can't identify the files from simple inspection.

JR

 

[JohnRoberts]

Care to elaborate on this? 

Abbey can speak for himself but "music" consists of a series of sound events that start and stop, not continuous signals that lend themselves to perfect processing.

Most formal definitions for continuous-time signals, like:
and:Audio (at least as a representation of sound pressure traveling through air) would seem to fit these definitions, at least between some sensible range of input values.

Mathematical (like Fourier) analysis of non periodic signals involve assumptions and errors.

This is getting off into the weeds about a simple listening test, but fair game in the brewery.

JR

 

[abbey road d enfer]

Care to elaborate on this?  Most formal definitions for continuous-time signals, like:
and:Audio (at least as a representation of sound pressure traveling through air) would seem to fit these definitions, at least between some sensible range of input values.

By definition, a continuous signal just cannot exist, it's a mathematical entity. All the theorems and formulae that constitute the base of digital audio are unattainable practically. There's no doubt a sinewave can be reconstructed with only 1bit quantization but it will take sometime to make it well identifiable considering half the samples may have 100% error. Will it be acceptable for practical recording purposes? I doubt it.

Quote: "at least between some sensible range of input values"
That's the heart of the subject. Knowing we are applying math that are meant to work with perfect idealized signals, how far is the result when we apply them to real-life signals? That's what determine the "sensible range of input values". In math, it's called the domain of validity.

 

[ruffrecords]

Let's not forget that audio is not a continuous signal.

In the spirit of christmas past, Oh yes it is.

Digital data is not continuous but analogue is.

Unless you are going to invoke quantisation in which case nothing is continuous.

The subject is made a little more complex by the fact that most converters are Delta/Sigma, which again involves dealing with strings of data that are interrelated.

Single Speed (44.4 and 48k SR) is inadequate, mostly because of the constraints on the anti-alias and recovery filters; Double Speed is more than adequate (in fact, if they had a time machine, many designers would chose 60-64k). Quad Speed has no operational advantage, though it is of some use to those who do measurements.

I thought oversampling eliminated the filter problems?

Cheers

Ian

 

[scott2000]

So this test is to decide whether or not we can hear differences in a higher rate and depth converted to a lower rate and depth?

It would be interesting if any differences are heard and even more if the truncated or dithered versions are more favorable....

I have to change my settings to play the files...lol....it's all slow motion......

Isn't "working" at higher rates where  things get more involved?

 

[Gold]

He should know better. A totally disingenuous test. Full mixes at final level will be hard to distinguish. A solo violin with a huge dynamic range  won't when it is ppp.

 

[emrr]

It does not at all address the reality of multiple conversions when mixing on an analog console, and recapturing multitrack digital as 2 track digital.  I'll use the large format camera for that, and reduce the final small print version when the time comes, thank you.  Nor does it address plug-in processing quality, which is tied to a sample rate.  Or specific converter quality at various rates.  Or playback system quality and resolution, and the room that's done in.  You cannot fairly simplify this test without a lot of caveats....such as 'this test is only about sample rate conversion and it's effects, or lack thereof'. 

 

[JohnRoberts]

In the spirit of christmas past, Oh yes it is.

Digital data is not continuous but analogue is.

methinks you are sawing on the wrong old log...

Yes digital samples are discrete (I hope), but the output of DACs pass through anti-imaging filters that smooth the waveforms.  Even these smoothed waveform are not continuous (for all time), but truly analog waveforms.

Unless you are going to invoke quantisation in which case nothing is continuous.
I thought oversampling eliminated the filter problems?

Cheers

Ian

This thread is to notify about a new listening test, not to revisit formal digital theory.

Unfortunately that is how the internet works.

JR

 

[JohnRoberts]

It does not at all address the reality of multiple conversions when mixing on an analog console, and recapturing multitrack digital as 2 track digital.  I'll use the large format camera for that, and reduce the final small print version when the time comes, thank you.  Nor does it address plug-in processing quality, which is tied to a sample rate.  Or specific converter quality at various rates.  Or playback system quality and resolution, and the room that's done in.  You cannot fairly simplify this test without a lot of caveats....such as 'this test is only about sample rate conversion and it's effects, or lack thereof'.

Indeed it doesn't, this addresses the playback media format.

Historically recording studios enjoyed significantly larger dynamic range and linearity than consumer playback media.  The introduction of the CD upset that apple cart as early professional digital systems had trouble keeping up with the then still unrefined CDs.

If I could hear a difference in my studio I would use the superior media for tracking (but I do not record ).

If consumers can not hear a difference, and even worse refuse to pay for a better medium (on paper or otherwise) the customer is always right, and gets what they are willing to pay for. 

Over the decades I have followed many arguments about what people say they can and can't hear on the WWW. I decided long ago to not argue  with people about things that can not be proved (easily).

This listening trial seem interesting but let's wait until it occurs to debunk it. I expect Ethan will be receptive to criticism of the methodology after it is done. It is probably inconvenient to stop everything and start over at this point. 

Plenty of drama at Ethan's facebook page where he announced this, but Ethan is no stranger to drama..

JR

 

[abbey road d enfer]

In the spirit of christmas past, Oh yes it is.

Digital data is not continuous but analogue is.

Unless you are going to invoke quantisation in which case nothing is continuous.

OK, wrong wording. I should have said steady-state or "permanent regime", which music is not.

I thought oversampling eliminated the filter problems?

It makes filter design easier, but still digital filters are not perfect (rounding errors). The resulting errors are more critical at Single Speed than at Double Speed, if only because the energy close to the Nyquist frequency is lower at DS, due to restricted frequency response of the analog recording chain and statistical signal density.

BTW, I consistently chose the lower resolution files as sounding better!

 

[ruffrecords]

BTW, I consistently chose the lower resolution files as sounding better!

Interesting, back in the 70s when I was a lot younger and my ears were a lot better, I was used to listening to tracks recorded on analogue machines that were essentially flat out to 25KHz. A decade later when digital came along I could definitely hear the difference between the original and a 20KHz band limited version. I am not saying I could hear up to 25KHz but I could tell the difference between material replayed in a 25KHz channel bandwidth and the same material replayed in a 20KHz bandwidth. Of course, the filters in those days were not as good and today my ears are not good enough today to tell the difference anyway.

Cheers

Ian

 

[john12ax7]

I pretty much agree with abbey on this one with the small exception of bit depth.  I hear a slight difference between dithered 16 bit and 24 bit.  For me something like 20 bit 64 kHz would be an ideal release format.

 

[living sounds]

With this test you're also judging the conversion quality of Sound Forge. I have only found a single conversion software I like, finalCD (freeware, offline, takes a lot of time and processing power).

 

[abbey road d enfer]

Interesting, back in the 70s when I was a lot younger and my ears were a lot better, I was used to listening to tracks recorded on analogue machines that were essentially flat out to 25KHz. A decade later when digital came along I could definitely hear the difference between the original and a 20KHz band limited version. I am not saying I could hear up to 25KHz but I could tell the difference between material replayed in a 25KHz channel bandwidth and the same material replayed in a 20KHz bandwidth. Of course, the filters in those days were not as good and today my ears are not good enough today to tell the difference anyway.

Cheers

Ian

I would not attribute the differences I heard to the different HF response extension. I just felt the files I preferred were "cleaner", whatever it means...

 

[abbey road d enfer]

I pretty much agree with abbey on this one with the small exception of bit depth.  I hear a slight difference between dithered 16 bit and 24 bit.  For me something like 20 bit 64 kHz would be an ideal release format.

If I was producing classical music or sophisticated jazz, I may wish for higher resolution than 16 bits, but as I do only rock and blues...