Upsample or Downsample?


Help Support QuadraphonicQuad:

This site may earn a commission from merchant affiliate links, including eBay, Amazon, and others.


800 Club - QQ All-Star
QQ Supporter
Since 2002/2003
Mar 11, 2002
Here's a situation that will probably generate diverse opinions ....

Suppose you have a client that needs a particular file (song or album) that must be at 24 bit / 48k resolution. You have two versions of the file. Both files are from the same high quality source (original master tapes, SHM, MFSL, K2 ,blah, blah, blah...whatever) but you do not know the bit or sample rate it was originally recorded. One is a 16 bit 44.1khz file and the other is a 24 bit 96khz file.

Question: Is it better to upsample the 16/44.1 file to 24/48 OR downsample the 24/96 file to 24/48? Which way would give the best sounding, most accurate results with the least amount artifacts or distortion?

My initial thought would be to downsample the 24/96 file - You always want to work from the best quality original and you are staying with the same bit depth - right? And the 16/44 file can not sound better simply by upsampling the bit depth and sample rate - right? But if you do a bit of searching on the net, there are many articles that argue that higher sample rates don't necessarily translate into better sound and some even argue there is less accuracy and more chances for distortion or artifacts at higher bit rates. Some of those same "experts" argue humans can't hear a difference between 16 bit and 24 (or 32 float) bit. It is also suggested that downsampling from a higher to lower rate can introduce distortion and artifacts that were not in the original recording.

For example: (That article addresses higher sampling rates as a native format, not in terms of upsampling existing files.)

To add to the controversy, you see people claiming many of the titles on sites like HDTracks are not really native hi-res tracks, just upsampled to a higher rate to make them "hi-res".

So, what do you all think? Is it better to upsample or downsample to arrive at the desired bit depth and rate of 24/48? (And would your answer be the same if your available source files were 24/88, 24/176 or 24/192?) :lookaround
Down-sample by decimation to 48kHz from 96kHz or 192kHz (i.e. even integer divisions) and truncate the results to 24-bits (you actually will gain a small amount of resolution!). In essence its how the single bit DSD systems work (single bit sampling at a very high frequency), although they do use a digital technique called noise shaping). It is possible to down-sample with non-integer division samples rates i.e. from 88kHz to 48kHz, but that requires interpolation & decimation. If down-sampling adds in artifacts and 'distortion' its not being done properly.

Up-sampling requires you to interpolate, so adding in calculated samples, some interpolation methods are better than others, and it is possible if done incorrectly to cause problems.

There are good technical reasons for oversampling, can the average person hear the difference? possibly not.

Sampling an analogue signal produces the same artifacts as mixing in an AM/FM radio, i.e. sum & difference frequencies around the sample (re: mixing frequency), hence why the input low pass filter should reduce the level of the input signal to below the resolution of the analogue-to-digital convertor (ADC) by half the sample frequency (ref: Nyquist), or the frequencies above this will be folded back down the spectrum and be audible. Which is one reason why over sampling is a good thing. Another is that it allows the use of 'smoother/less aggressive' filters, Bessel filters for instance have a less sharp roll-off above the cut-off than a Butterworth (often used) but have a very good minimally damped transient response. Listening tests have shown that some (not all) people prefer/notice slower/smoother roll-offs, its possibly also why some people don't like the 'brick-wall' spectrum cut-offs found. Like wise up-sampling digitally (i.e. interpolating) produces unwanted frequency components so the signals have to be digitally filtered.

Convertor resolution has to be accurate (i.e. linear across the input range) or it will introduce distortion, the majority of current 24-bit ADCs and DACs are good at 192kHz sample rates.
I'd agree with downsampling. I'll use data theory as my first justification. Information cannot be added to data without some other information being included (such as a smart upsampling algorithm). Simple upsampling will not not add information. But, given you are starting at 96/24, then 48/24 is still more information than the 44.1/16 upsampled can provide.

Second argument is that the people who argue that 24 bits versus 16 bits cannot be heard have not accounted for the low level instruments only using a fraction of the 16 bits. Not everything is sampled at a full 16 bit word length range. If you separate out the low level instruments you can hear how distorted those become by only using a 16 bit word length. So what a 24 bit word length provides you is the ability to have more bits for those low level instruments. Just because the word length is 16 bits doesn't mean that every part of the music takes advantage of all 16 bits. Those low level signals could be using only 6 bits or less. It is one advantage for DSD since you don't have a word length to worry about.

I would have to agree that I can't hear a difference between 48-kHz and 96-kHz sampling rates.

My only suggestion is that whatever you do, you only do it once. Multiple conversions is a good way to introduce errors. And, as DuncanS said, only use integer conversions if possible.