Down-sample DXD 352 to 192 or 176, and 96 or 88???

[mansr]

Integer resampling ratios can be less computationally intensive to perform, but there's no difference in quality if done properly. I assume whatever software they used is decent.

 

The real question here is why you didn't downsample the files yourself rather than buying the album twice.

[mansr]

6 minutes ago, PAP said:

convenience, and same friend told me that the converters at the various labels are mostly better than my jriver...

I doubt you could hear any difference between jriver and whatever they use. I know some labels use Sox for conversions, and that's free.

[mansr]

3 minutes ago, PAP said:

Sox. Ok I will look into that, thanks!

But if you had the choice and were going to download one of those down samples would you choose 192 or 176,

or indeed do it yourself ... you would convert to 88 or 96?

If buying I'd pick the cheaper one, or the higher rate (just in case) if they were the same price. There's no reason the quality should differ.

[mansr]

4 minutes ago, elcorso said:

Several years ago there was a strong and long discussion in this forum regarding this, where several recording engineers and audio software developers participated. I can't find the thread !
The answers (as always) were controversial.

I personally (and a group of friends) prefer to respect the multiples. The same regarding our preferences toward Native DSD vs DoP.

Could you describe what you compared and the difference you perceived?

4 minutes ago, elcorso said:

At this time I try to download the version at 176 (the server is very busy ), before, I downloaded the 128 DSD. 

Unfortunately there is no information from Sound Liaison if the simultaneous recording on analog tape was used as a source for some of the formats.

You could ask them.

4 minutes ago, elcorso said:

PS/ Please note that is an opinion only, based on taste, and not to start an never ending discussion !!!

My opinion is based on maths.

[mansr]

1 hour ago, Kal Rubinson said:

Exactly, apparently some systems require significantly more CPU effort to up- or down-sample between non-integer multiples than between integer multiples.  With hig-rez and multichannels, the difference can choke the processor leading to interruptions in the playback.   The difference can also be measured with any number of tools.

 

However, when all runs smoothly, I do not hear a difference between non-integer and integer conversions.

The original question concerned offline resampling where CPU load is much less of an issue.

[mansr]

17 minutes ago, Fitzcaraldo215 said:

It just makes more mathematical sense to stick with the integer multiple idea,

No, it doesn't.

[mansr]

10 hours ago, dtc said:

The idea that integer downsampling retains many of the original data points is why people think integer downsampling must be better than non integer down sampling.  In fact, modern methods use detailed interpolation algorithms that are far more accurate than just dropping intermediate data points.. They use the trend, not just the individual points.  You really have to look at the actual algorithms to be sure, but most people  who know the algorithms say integer downsampling is no more accurate than non integer downsampling.

Simply dropping samples results in aliasing of the high frequencies into the remaining band. To avoid aliasing, a low-pass filter must be used. Integer downsampling can be regarded as a low-pass filter followed by dropping samples. A practical implementation is simply a low-pass filter with a short-cut to calculate only every N samples for a factor N downsampling.

 

Downsampling by a non-integer ratio is a little trickier since only some output samples coincide in time with input samples. A common method is a polyphase filter. Recall that the output of a filter is the convolution of the input with the impulse response of the filter. To calculate the output value between samples points of the input, we can simply perform the usual convolution with the impulse response evaluated at corresponding inter-sample positions.

 

As an example, suppose we wish to reduce the sample rate by a factor of 2/3. With an input sample period of T, we get an output period of 3T/2 and samples at times 0, 3T/2, 3T, and so on. Every second output sample coincides with every third input sample, while the rest fall in between. The first output sample is calculated by convolution of the input with the impulse response as usual. To obtain the second output sample, at 3T/2, we first evaluate the impulse response function at T/2, 3T/2, 5T/2, etc, then convolve the input with this sequence. The third output sample is aligned with an input sample, so we go back to the normal impulse response. We then carry on alternating between the two impulse response sequences for the remainder of the signal.

 

This approach works for any rational resampling ratio. In the example above, we needed only two impulse response sequences (or filter phases, hence the term polyphase) since the ratio is simple and every second output sample lines up with an input sample. More generally, an N/M ratio requires N filter phases. For example, resampling from 48 kHz to 44.1 kHz, a ratio of 147/160 needs 147 different convolution sequences. Integer downsampling is just a special case with N=1.

 

In practice, it is common to pre-calculate the impulse response for the required phases and store all the sequences in RAM. For very long filters, this can end up needing several megabytes of space. If this is larger than the CPU cache, it will impact the performance though it's difficult to predict how much.

 

As for accuracy, the actual calculations are essentially the same regardless of the resampling ratio. There is thus no reason for any ratio to produce more accurate results than another.

[mansr]

2 minutes ago, Fitzcaraldo215 said:

Thanks for the very thorough and knowledgable explanation.  

 

It seems the consensus and my own operational practice agree that, in case downsampling is necessary, it cannot hurt to use integer downsampling, and there is nothing to be gained by not using it in that case.  It probably does not make a big sonic difference either way, unless computer resources are extremely tight in on-the-fly conversion, where integer downsampling may be preferred over non-integer.  But, no downsampling may be best wherever possible.

 

Does that seem fair?

All else being equal, choosing integer ratios will indeed not hurt anything, so by all means do that if it makes you feel better.

 

Btw, the explanation above applies with only minor changes to upsampling as well.