Sorry for interrupting your conversation, but I don't want to create another topic only for this:
I would like to know if there is any difference in playing 16bit lossless file with 24bit or 16bit output format (i use foobar). Is there also a difference (degreed/improved/same quality) for lossy formats? And while my card support 24bit output, should I use it instead of 16bit.
Almost all my files are 16bit flacs, but I need some space so I'll rip my cds with mp3 until I get a new hdd.

An you should prove the playback chain is really 24 bit capable (ie not rely only on the specs of the hardware/software).

I'll try to answer this.

With the gliding tone examples at http://64.66.184.67/technical/bits44/index.htm on the PC ABX Web Site using foobar in ABX mode and set so as to play only the last few seconds of the samples (i.e. the quietest part), setting the foobar output at 24-bits seemed to give the best performance. What I found was as follows:-

The result with the option set to '24-bits' was good for sample 1.1.2 (16-bit dithered). The dithering algorithm used for preparing the sample must have been of good quality. Switching the output option to 16 bits gave a much noisier result (and this was much the same result whether or not dither was selected as an output option in the foobar preferences).

With sample 1.1.11 (16-bit, not dithered), artefacts were more noticeable (spurious whistles) on my setup (an Audigy 4 with external hub) than with sample 1.1.2.

Setting the foobar output to 24-bits gave the quietest output. Setting the output to 16-bits increased the noise (somewhat hiding the whistles). There was little difference between setting the foobar output dither to on rather than off.

With playing of 16-bit music more generally, I don't see why it should make much difference whether the codec is lossy or lossless. If it is lossless but derived from a source digital recording at more than 16 bits the source needs to be reduced to 16-bits resolution somehow using dither, rounding, or -- conceivably -- truncation. It has to be remembered that 16-bit 'lossless' still has the limitations of 16-bits: it is a perfect copy of a 16-bit resolution approximation. The question we are trying to answer in this thread is whether that approximation is so good that no difference can be detected between 16 bits and a greater number of bits when listening to music. I think it must be assumed that the gain used when listening to the test extract cannot be unrealistically high. By unrealistically high, I mean that the loudest parts of the music must not be unbearably loud when played back with the same gain setting used to listen to the test extract.

It is trivial to prove that 24-bit music sounds different to 16-bit music if the 24-bit source has a very high signal to noise ratio, part of the performance is extremely quiet, and that part of the performance is listened to at an unrealistically high gain setting.


Choosing a quiet part of a high signal to noise ratio 24-bit sample should enable the noise level to be assessed. When 16 bit sound-card mode is selected, background noise should rise quite substantially compared with operating the card in 24 bit mode. Foobar conveniently makes the output bit-depth selectable and also allows playback to be confined to just a selected part of a file.

Obviously this is only a rough and ready test, but it may be sufficient to indicate that the system is operating at more than 16 bits.

AndyH, perhaps you might be kind enough to answer my two questions below.

Being a newcomer to this site, I’ve only now read some of the other threads on the 24 vs 16 bit debate. I’ve also now read some of the material on the complex subject of dithering.

Before submitting any sample of 24-bit audio that contains a section that can be demonstrably distinguished from 16 bits (testing with ABX software), I’d just like to check on the ground rules for such an exercise.

Question 1:
Given that it can be demonstrated (using low level test tones) that artefacts in a 24-bit to 16-bit conversion are more evident if dithering is not used, and given that dithering is widely in use and recommended, is the use of some form of dithering compulsory for this exercise? If it is not compulsory, then I might seek to identify a sample where artefacts were in evidence.

Question 2:
If the answer to question 1 is that dithering is compulsory, or at least acceptable, and given that dithering introduces an amount of noise, does this exercise exclude from consideration simpler forms of dithering that introduce much more noise than is necessary? More specifically, are any of the dither methods available with readily available software such as Audacity or foobar acceptable for this exercise? If acceptable dithering method(s) can be indicated, I would seek to identify a sample where the noise of the dithering is detectable when comparing a derived 16-bit version with a 24-bit original.

************************************************************

What is the theoretical signal to noise level available for 16-bit format of a conventional audio compact disc?

It is often suggested that 16-bits provide around 96.3 dB of signal to noise ratio, based on there being 2^16 instantaneous levels that can be represented with 16 bits. However, it’s an extremely complicated question. The S/N ratio can be argued to be infinite (if a series of zeroes is used between tracks of the CD), or extremely high (if a series of zeroes is sent interspersed with an occasional positive bit, or occasional negative bit [which could apply is using high frequency dithering to make a low frequency low level tone audible]).

Substituting a slightly different (and much simpler!) question, for sine waves with no direct current offset and which are at a frequency not far below half the sampling rate, the maximum [peak] amplitude that can be represented either side of zero is half of 2^16 , and the minimum [peak] amplitude that can be represented either side of zero is 1. Half of 2^16 is 32768. Voltage ratios can be expressed in decibels using the expression 20 x log (voltage ratio). That gives us the figure 90.3 dB as the ratio between the peak level of the loudest sine wave and the softest [highly imperfect] representation of that sine wave without dither. The low level wave doesn’t really have a shape to make the description ‘peak’ meaningful, being some variety of rounded square wave. And the filtering needed might alter the peak value of the low level ‘sine wave’ somewhat.

Recording at 24-bits and keeping the noise below that which applies for the 16-bit format

Without attempting to pin down a definitive answer to the S/N ratio question, I have found in practice that the required signal to noise ratio of recording equipment needed to compete with that available in the 16-bit format is quite high. What I have found is that is stretches my equipment to the limit. A Behringer analogue mixer struggles to provide a signal that has noise much less than 90dB down when set up so as not to clip in loud sections of a recording that approach 0 dB. A Creative Audigy 4 soundcard with external hub struggles to provide better than about 94 dB with some audio drivers. By locating the Audigy 4 card in the pc in a PCI slot away from interference, and by using an ASIO ALL driver (instead of the normal a creative driver), and by setting the electronic fader in the creative recording mixer to less than 100%, I was able to get over 100 dB available for use for 24-bit recording.

Then there is a question of room noise where the recording is made. A close microphone technique is needed.

When all of this is done, I find that a dithered 16-bit version of a 24-bit recording can be audibly noisier at a realistic (though rather high) listening level, during quieter passages of a recording.

However, a lot may depend of the type of dither used. I would prefer not to submit any sound recording extract until I have a better understanding of what forms of dither are considered acceptable.

Yes, dithering should be compulsory for this kind of test. It is well understood than the quantization process works "better" if there is the "right" amount of noise added to the signal. This has to do with reducing the correlation between the quantization error (or residue) and the signal to be quantized. Where error is not correlated with the signal, the only result of quantization is the addition of noise - not the addition of distortion.

Not applying the optimal type of dither before performing such a test would seriously call into doubt the validity of the results.

If your hypothesis is "reduction of 24bit music samples to 16 bits using optimal techniques is not transparent", then you would have to use the theoretically "best" dither available. If your hypothesis is "reduction of 24bit music samples to 16 bits using X technique is not transparent", then you can use any kind of dither available. Note that confirmation of the second hypothesis is still useful.

************************************************************

Without dither, some of the assumptions behind the quantization process break down. In most basic analyses, it is assumed that the quantization error for high bit counts is not correlated with the signal to be quantized - where this does not hold the analysis becomes much, much more complex. I would recommend borrowing or buying a copy of a good DSP book (Oppenheim and Schafer, for example) for answers to these questions better than those that I can offer you.

Hi!

I'd like to remind you that there's a free tool around for requantization, dithering and noise shaping that supports arbitrary bit depths and custom noise shaping filters: NOISE

Note: The built-in filters as well as the other supplied filters might not be the best ones. I encourage you to try creating your own filters.

Using the appropriate noise shaping & dither settings 12 bits/sample @ 44kHz already sounds good enough to me. So, I feel quite safe with the common 16/44 format.

Cheers,
SG

I know this isn't the only time I raised the topic in this forum, and I have done the same thing in a number of other forums. Every time I write sort of off the cuff, that is, without a preset script. I'm not in the mood to reread this entire thread so I can't say exactly what I might have already discussed in detail here, as compared with other times I addressed the subject.

However, I did include the necessary basic conditions in the first post:
a properly resample to 16 bit version of same

It may be possible, using poor quality dithering, with no noise shaping or poor noise shaping, to make an audible difference because of the added noise. That is not "properly" done. Many people here will remember that in the recently published year long ABX test of DVD-A & SACD vs resampled to 16 bit, the only differences detected by anyone were the "silence" between tracks, turned up to extremely high levels, where the unshaped dither used in their resampling was audible.

Proper resampling means, to me, what CoolEdit/Audition does with resampling, dithering, and noise shaping. I won't go into the evidence that has been presented in various places, but the program's resampling is probably as good as it gets.

It also has a fair number of dithering and noise shaping options. They might not be equivalent to the industry's best, but some of them work extremely well. I am only interested in the options that work well. The only point of using something else would be to encourage failure, not something anyone knowledgeable would do with a real music project.

So, I'm looking for a 24 bit music sample that I can resample to 16 bit in CoolEdit and tell the difference. I know there is other very good resampling software, but I also know that quite a few of the commonly used mixing/mastering programs do a relatively poor job of it. Just saying that someone has managed an ABX test isn't satisfactory without adequate evidence that the resampling wasn't done in such a way as to encourage an audible difference.

Cool tool, thanks for letting us know about it (and writing it )



Good post, but you neglect the fact that there is an important difference between changing the sampling rate (resampling) and changing the number of bits per sample. In my opinion, the latter should never be called resampling, as this is misleading.

Thanks cabbagerat for your various remarks. I am not sure I am all that attracted to the idea of proving that last hypothesis, 'X technique is not transparent' if better dithering is apparently so readily available. As I have recently discovered, dither using Audacity appears to be much better [certainly quieter] than another method I've used (software I paid for!), and Audacity is a free download.


Thanks, AndyH. Just as well I asked. I had assumed the dither that is built in to some software I've been using for a few years, n-track studio, would do a passable job in terms of added noise. But I find that is not the case, now that I've started comparing the results of different forms of dither.

With some recordings I have recently made at 24-bits, and processed to 16-bits, I have obtained these results:

1. N-track Studio 4 set to 1-bit dither - the added noise is quite noticeable.

2. N-track Studio 4 set to 1-bit dither plus shaping - the added noise not as noticeable as in 1, but nevertheless is detectable at a loud but realistic listening level (detectability verified with ABX software).

3. Audacity: dither left at default (triangular) - the noise is slightly different to that in the 24-bit original but so low in amplitude I think it unlikely to be detectable at a realistic listening level [by me anyway].

4. Audacity: dither set to shaping - same as 3, though a slightly different quality to the noise.

The extracts in which I have detected differences due to the noise level have used the dithering method of 2 above, so I guess that will not be good enough for the current exercise.

If I am to pursue this further, I will probably need to concentrate on other artefacts than noise, though the more I read, the less optimisitic I am of success.

It makes current complaints, in other web forums, about home theatre personal computers not being able to access all 24 bits from some Blu-ray audio formats (partly because of unresolved issued regarding HDMI and content protection protocols) seem unnecessarily alarmist, unless there is no dither used for the reduction to 16 bits. It even has to be queried why 24 bits are being transferred in the first place to some Blu-ray discs, which I guess is what was behind the DVD-A & SACD testing you mention. In relation to surround sound, I guess where 6 or so full range channels are in place, the total noise floor may begin to be an issue, particularly if a movie is created with a very wide dynamic range. It may be a slightly different exercise to produce a two channel CD at a normalized maximum recorded level.

When I first used a 24-bit audio card, a few years ago, for capturing amateur musical performances, I could immediately detect the improvement compared with capturing with a 16-bit sound card. However, it appears that provided dither is used in the mixdown from a 24-bit source, enough of the 24-bit intensity detail should normally find its way into the 16 bits, at least at lower signal frequencies, where we might otherwise detect the lack of intensity resolution.

I haven't given up, but as I say, I am no longer as optimistic.

Not to be pedantic - but the only objective conclusion you can draw is that you immediately noticed a difference between the quality of the two cards. Unless you performed more testing than you have indicated you have no reason to believe the quality difference you detected had anything to do with one being 16 bit and the other 24 bit.

Thanks, Soap. I should have been more precise. When I started using the new card (an Audigy 2), I compared recordings made with it with its drivers set to 16-bits and with drivers set to 24-bits; and the 24-bit setting gave a noticeably different result (not a dramatic difference of course). As I was not intending to release the recordings in red book CD format, or other uncompressed 16-bit format, I never investigated dithering techniques, though I was aware they were recommended.

I'd also mention that once I had a 24-bit card available I no longer sought to make recordings at close to clipping level, but allowed the recordings much more headroom.

With proper dither, there will be no artifacts other than noise.

I don't think anybody really doubts whether 24 bit is better than 16 bit for recording, mixing, processing, mangling, mastering, etc. the sensible debate is only whether 16bits is enough for final delivery to the customer.

If I am not mistaken, that is the question we are trying to resolve, whether the arte[/i]facts that can be heard (even with optimal dither) when test tones are used, set at very low levels, and listened to at a high gain setting, can also be heard by the human ear at realistic listening levels, with music.

I found the following software quite useful for getting a taste of artefacts:-

QUOTE(jlohl @ Sep 6 2007, 20:02)

QUOTE

What is the effect of a non-white random source in dithering?

I did a small and free software to show/hear the effect of dither. The dither itself can be noise but also a sine wave. So you can try and listen to the result.
Get it here

Well, if that's the question, then to my mind it's already been resolved for 24bit -> 16bit conversions. Consider our 24bit music samples x[n] and dither signal d[n]. Now, we define e[n], which is the difference between the x[n] and quantize_to_16bits(x[n] + d[n]). Ok? Now, optimal dither would do two things:

1) Make e[n] and x[n] completely uncorrelated, and
2) Reduce the total power of e[n], while still achieving (1), or
3) Reduce the audibility of e[n] while still achieving (1).

From what I understand, for the case of quantization from 24bits (or 32 bits) to 16 bits, the formulation of the jitter signal d[n] to achieve (1) is well understood. The formulation of d[n] to achieve (2) and (3) is a little bit less well understood, but still can be very well approximation with common processes. What (1) means is that there is no distortion at all caused by the quantization process - only the addition of noise.

I think the whole 24 bit/16 bit thing boils down to the hypothesis:
Consumer music and movie soundtracks have sufficient dynamic range, that when listened to at "normal" levels, the noise introduced by optimal dithering and quantizing to 16 bits is audible.
In my opinion, this hypothesis is yet to be resolved conclusively for real world music and movie soundtracks. This has been demonstrated to be true for some classes of test signals.

Er, yes, "ok" after several re-readings!

Expressed non-mathematically, I can understand how a high frequency dither can be used to supplement missing intensity level detail for lower frequency signals; and to mask the remaining low levels of quantisation error.

I've decided that my chances of being able to hear artefacts when playing back real life music in a 16-bit format derived from a higher bit depth source, and using a quiet dither, are likely to be confined to the artefact of the dither noise itself.

Using a 300Hz test tone at -60dB, my AVRs when fed a 24-bit PCM bitstream, play the tone with some background noise, and a small edge of distortion. Substituting a 16-bit PCM undithered version of the test tone, the distortion becomes quite pronounced. Adding triangular dither, the pronounced distortion disappears but dither noise is quite apparent. [It is certainly the lesser of the two evils.]

Minus 60 decibels is a very low signal compared with 0dB, a ratio of 1:1000 in fact. And the dither noise is quite a bit quieter than that.

The noise level of my microphone preamplifier/mixer competes with the dither noise if the gain is set to avoid clipping at 0dB. So to make a recording of music (e.g. a pianoforte) where the dither noise was appreciably greater than other noises, and to include in the recording a passage within a few decibels of 0dB, I would need a new preamplifier. [Fortunately the microphones have very low noise.]

If I were to succeed in this exercise [ABX tested, and dither method acceptable], what would I have proven? Simply that at very high (many would say artificially high) listening levels, dither noise can be audible when listening to a high signal to noise ratio 24-bit stereo recording of music, reduced to 16 bits.

That is not the narrow result I had assumed. I had assumed that some other artefacts might be audible. But I gather that other artefacts will almost certainly be neutralised, or sufficiently masked, by the dither noise.

I generally issue this type of challenge when someone is loudly proclaiming the obvious superiority of “high definition” audio. That generally includes both bit dept greater than 16 and sampling rate greater than 44.1kHz. I don’t recall just why I open a new thread to ask about 24 bit, maybe just to see if anyone could come up with something in a non-confrontation posting (every time I’ve asked someone who is so certain of his/her beliefs for evidence, the conversation has stopped). I don’t think there is an way to differentiate except by way of the dither noise, and rarely even then, but the question is genuine. If such music exists, I would like to experience it.

I suspect, however, that even if someone can come up with a real sample, it will only be noticeable in a close comparison of the two versions. If one is just listening to a performance, one could never figure out which it was.

I've a added a 24-bit 48KHz original, and 16-bit dithered derived version, in the uploads forum under a new topic Short samples: 24 bit vs rendered to 16 bits dithered. It's 9 seconds of a simple piece by Handel, played on an upright piano by an amateur pianist.

These wave files certainly show a difference in noise level as a result of the dither (if listened to at very high gain). They also seem to show a difference in the sound quality when listened to, briefly, at high listening levels. However, before making too many further comments, I await any critique of the dithering method.

I presume comments on these files may be more likely to be made in the uploads forum under the new topic, rather than here, but I'm not sure what the preferred practice is.