24 bit is definitely the way to go if you properly dither before converting to 16 bit; with this method, you can get performance that sounds like 18 or 19 bit due to the fact that signal can be perceived in the presence of noise (the dither) even when the noise is up to 12 dB higher than the signal. I am of the opinion that you should sick with 44.1. The math involved in converting 96 to 44.1 is messy. If you must go higher, use 88.2 since the SRC is much simpler.

David French wrote: The math involved in converting 96 to 44.1 is messy. If you must go higher, use 88.2 since the SRC is much simpler.

I agree both on 44.1 and 24. If nothing else, that is what I use myself all the time.

But for the statement above, it sound like part of the internet myths. So I simply ask two questions and as I keep the text to a minimum, please, I am not trying to be rude.

1 - are you sure? Have you really written any SRC routines yourself or are you only guessing.
2 - have you listened to SRC done in a modern welldone application and actually heard the 88.2 sounding better than 96 after SRC?

Gunnar

Excellent and fair questions Gunnar.

I have neither written routines nor heard audible differences, but I do understand the mathematical process that produces SRC, and it is pretty complicated going from 96 to 44.1 since there is no easy common factor. I am only guessing that this complication would result in an audible loss of fidelity do to things like rouding errors.

I don't pretend to know the math either, but I've heard it from those that DO to support this theory...somewhat. I have also heard that 88 is probably better than 96 to go down to 44, but more importantly, I've heard that the word length/bit depth is much more critical anyway. (24 bit vs. 16).

For most of my work, I stick with 24/44, and stay in that realm until it's time to make a CD. (I also bounce multitrack mixes down to 24/44 and don't dither until it's time to make the final CD version. DIther once, dither LAST). Rarely does a client even ask to go to 96k. (Honestly, most either don't hear a difference, nor do they care or understand the concept.) When a project is complete, I save everything: the multitracks, the 24bit mixes, and of course the CD master (16/44).

For various SRC situations, I remember reading once that it's better to go UP (from 44) to a higher resolution (like 96) before going back down to 48, or vice versa. There are also those that feel it's just as good to go out via ANALOG patching and just convert it that way. (whatever works, I say! ;-)

I don't want to cloud the waters further here with hearsay and fuzzy math; it's a good subject to research and discuss on your own, if you want. Personally, I just try to avoid SRC/gearboxing from things like 48 to 44 or 44 to 48, for example, unless I'm very sure of the software/converter that's doing the job at hand. That just seems to make a lot of sense to me.

I always try to ascertain ahead of time where the audio is going to end up: on a CD, Video, DVD, broadcast, etc. and then do the best job possible for the intended format. (Having said that, how many of us have grabbed 16/44 wav files and dumped them into the timeline of Sony Vegas or Adobe Premier, used those program's built-in SRCs, crossed our fingers, and hoped for the best when it came time to render the final soundtrack? ;-) Again, I try to do this on my own, with better software, and create 48k versions of the files BEFORE moving them into the video timeline, or start the project at 48k in the first place. )

Fortunately, with my 24/44 2-mix (Stereo) masters, I can usually take it any direction thereafter without too much tradeoff - be it a CD, video sound track, WAV file for broadcast, etc.

Of course, we could also talk about why 24 bits is something of a myth, and why it's really NOT 24 bits, eh? :-)

I agree again. When possible I think it is better to avoid the SRC.

I must admit that I have never written an SRC either. But I think I know how I would go about it in order to make a good one. (Mind you, from a theoretical point of view). It takes a lot of processing power when you select the "ridicously slow and very good algorithm". Only a few years ago CPU power was too limited to even offer that way. So in those days you had to "cheat".

So from a theoretical point of view there is no difference in going from, say, 88.2 to 44.1 as from 96 to 44.1. But the "cheating" may have effects in various ways, and here is where you have to use your ears, they may sound different.

As for 24 bit, well, all the bits are there, that is no myth. It is only that no A/D converter in audio gear I know can resolve all the 24 bits with information. Most end up about 120dB S/N (sound to noise) and that is more like 20 bits.

Gunnar

88.2 to 44.1 produces no aliasing. The SRC simply throws out every other sample. This is integer-ratio (synchronous) SRC, where the target rate is always coincident (a direct multiple) with the original rate.

96 to 44.1 is variable-ratio ASRC (asynchronous, where there is no fixed fractional relationship.) In this case the SRC oversamples the original rate to a direct multiple of the target rate, so as to then throw out (decimate) the in between samples.

Some say that even the high quality filters involved with the ASRC method can't remove all the sideband temporal artifacts caused by interpolation...making this the less transparent of the two.

Others say the two methods are equally transparent.

I've heard both (once, mind you); & my ears went with the equal math.

Having established that...the samples which are retained from the original higher rate (even in their fractional re-compostion) still contain the upper harmonics & spacial overtones recorded at the higher rate, right? And even though we can't hear them...they still contribute to the sonic realism of the signal, right?

This part of the debate is older than the "conversion between varying rates" question (& more intense.) Can our brains perceive it...when our ears can't hear it?

Anyway, getting back to the reality of the question...it may be a mute point. 96 to 44.1 on a high-end system does sound better than original 44.1. Cheap SRC sucks...better to stay at 44.1. 88.2 to 44.1, even recorded into a project studio DAW & converted in a software audio editor (WaveLab), sounds a bit better (a little brighter) than original 44.1, to me.

my 2 cents,

mark4man

BTW - There's an audiophile test CD floating around somewhere with various conversions between differing rates...done on, I believe, a Weiss SFC2. If anyone's interested, I'll see if I can locate it.

Mark4man,
once more. Do you really know that your description, throwing away every second sample, is how downsampling from 88.2 to 44.1 is done? In any modern application? Can you name that application and a reference describing that this is how it is done?

Or is it once more the internet myths?

I know enough to say that trowing away every second sample is even worse than cheating, it is downrigth destroying the sound. I would never even suggest that kind of algorithm, as I would expect it to sound very bad. Perhaps they had to do it that way in the very early days of computing, but hardly today. But if that was the way it was done, I would expect 88.2 to 44.1 to sound very worse then 96 to 44.1. Go figure.

One thing you have to take care of in a SRC is filtering away the sonic content that cannot be fitted into the new sample rate. This means, as example, with 44.1kHz sample freq all content above 22.5kHz (and perhaps a bit of a margin as well). This can be done, and the algorithms for doing that can be found in standard text books used on university level.

I still keep pointing out that theory is one thing but you still have to listen.

Gunnar.

Gunnar,

What the h_ _l are you taking about...internet myths?

If the SRC is designed as a Synchronous SRC, after low-pass filtering (which is I think what you were getting at in your last paragraph), the odd samples are decimated (for integer-ratio SRC, e.g., 88.2 > 44.1.)

What is this... a quiz or a contest?

Gimme a break,

mark4man

Gunnar,

With all due respect, Mark4man is absolutely correct. And if you look in the mastering forum under "48K or 44.1?," you'll see that Ed Littman and I had a rather interesting debate on the merits of various types of SRC. And as a person who works in a shop where Sample Rate Conversion algorithms are written all the time, I can tell you that, yes, in a conversion from 88.2 to 44.1, the odd samples are discarded.

Bear in mind, this has nothing to do with filtering the recording of all information above 22.5 kHz. This is done with a steep low-pass filter and not by the removal of samples. And since all notes produce overtones both higher AND lower than their stated frequency, there will still be much of the information retained from the higher sample rate recording. This is fact, not myth. True, it may be difficult to hear, and in some (rare) cases, detrimental to the sound. However, it is scientific fact.

I've placed a simplified version of the sample rate conversion concept below - this is directly quoted from the discussion found on the mastering board.

Sample Rate Conversion is the state of taking 48,000 samples per second and transforming it to 44,100 samples per second. The only way to do this is to selectively remove some of the samples (nearly 10%). However, this cannot be done simply by randomly removing samples (despite the fact that this is what some Sample Rate Conversion programs do.) One has to be very selective at removing samples.

For example, you can remove more samples at lower frequencies than you can at higher frequencies. However, despite this, you simply cannot remove samples in an exponentially repeating pattern; the human ear recognizes this pattern. Therefore, you must do kind of a combination of the both (with a little bit of planning behind it.)

Removing odd or even samples doesn't destroy the sound. What makes you think this? Imagine a 2400 dpi photo - now, take that photo (with VERY small dots) and remove the odd dots. Considering the size of the dots, your eye would never detect the difference. Now, multiply this example by 20,000 - you will quickly realize that there will be no discernable change in the sound.

A more "scientific" way of thinking about it - to chart a frequency, you only need two points on that frequency - this will tell you the width and height based on various assumptions and equations. Until you get up to 20,000 Hz, you will only need 40,000 samples. If you have a sample rate of 88,200, you will have over 4 samples for each frequency at 20,000 Hz. Meaning, you can remove 2 of the samples out of the 4 without any detriment to the sound. You would want to remove every other sample, or you will wind up with errors in which the wave cannot be plotted based on misplaced samples. The low-pass filter gets rid of these higher frequencies that are now too high for the SRConverter to plot and then plots the rest.

Hope this helps,
j...

Ok. I read your posts. I agree with you, once I understand where you come from (sorry, I am bit new in this forum area, I looked not on the posters history, only on the post as such).

And mark4man has changed his description, or somehow I missed that in his first post.

The important point is that there is a bunch of people who has no idea at all about how a SRC is done. They think you simply remove every second sample going from 88.2 to 44.1. And hence it must be much more complicated to go from 96 to 44.1, any old fool can see that? And since it much easier it has to sound better! This is pure internet myth. And I tell you, there is people out there that can kill in holding up that myth.

Well, we both now that is not the FULL story. You have to do a low pass filtering somewhere. There is sonic content between 20 and 40 kHz in the 88.2 signal, and you have to filter that away. Otherwise it will be folded down into the audible frequencys. So you cannot simply throw away every odd sample, you have to do a bit more. And doing this does take a bit of processor power, and I believe that here is where the "cheating" is done in some algorithms.

So if anyone suggest that SRC from 88.2 to 44.1 is ONLY to remove every second sample, I get a bit aggrevated. Could you unerstand that?

And, secondly, do you agree that listening is the only way to evaluate how things sound? (Well, perhaps apart from the really sour lemons then).

Thank you for the pointer to the mastering forum, I will go there.

Gunnar.