Hmmm, more than a couple of questions here.

First of all, keep the sampling rate and wordlength issues completely separate. They are orthogonal, and conflating them just confuses things. Secondly, I assume you are talking about using digital signal processing methods on the original files and not reproducing the signals in an analog form and re-recording them at the new sampling rate. Thirdly, apart from your last section concerning DSD-encoded data, I assume this is all PCM data you are concerned with.

When performing SRC (sample-rate conversion) to generate a direct sub-multiple of the higher sampling rate, things are a little easier from the clocking point of view, but the basic principle is the same as non-multiple. You have to use a method that in essence low-pass filters the original data (using the original clock rate) and then decimates the output data, i.e. takes 1-in-n points, discarding the rest. There is then a level adjustment to compensate for energy loss.

Sample-rate change algorithms do not change the dynamic range of the material or in themsleves introduce distortion. Down-sampling can actually improve the signal-to-noise ratio, as more of the acoustic energy is in the pass-band of the down-sampling filter. This is the principle used in over-sampling ADCs, where a sampling decision process is chosen that pushes noise into the high frequency band where it is removed by the LP filters prior to decimation.

The big problem with SRC is the quality of the algorithms themselves, which includes the type of digital filter used. I'm sure there are good ones, but for audio work, I often re-sample a high-rate analog output at the new (lower) rate. Despite the small increase in noise and distortion this introduces, I generally prefer the results to any of the SRC algorithms that I have available for use.

Turning for a moment to wordlength reduction, this should be done by conventional dithering algorithms applied to the down-sampled signal. This is so that the SRC process has the maximum wordlength to work on. So, in your case, once you have your 100KHz 24-bit signal, dither and truncate to 16 bits. Note that there are considerable differences between dithering algorithms that I won't go into here, but it's necessary to choose one that gives good results under your conditions. From the numbers you give, this is not conventional audio, so recommendations of dithering types that work well for generating domestic CDs may not be the best in your case.

As for DSD, we have had some threads on that topic in this forum (for example, [[url=http://[/URL]="http://recording.or…"]this one[/]="http://recording.or…"]this one[/]), and the basic problem seems to be the difficulty and cost of editing DSD-encoded audio. If you are concerned mainly with rate reduction rather than splice edits, this may not be a problem for you. Again, the internals of many ADC bricks show the sort of thing that can be done, and there are several research papers around showing methods of rate reduction and conversion of DSD to parallel-encoded data.

It's difficult to give more specific information without further details of your material, but I can understand if you require to keep that aspect of it confidential.

Hi Boswell,

Thank you very much for replying. Reading your reply I realize there is more to it than I had assumed - so I really appreciate your feedback which clarifies matters sufficiently for me to know which path to take in my design considerations.

And then a comment on what I'm working on as part of it is not really confident: It is an application mainly for audio but also related to measuring "energies" (if interested see e.g. http://www.item-bioenergy.com/rfi/RFITechnicalManual.pdf - chapters 1 and 3 may be a start).

Greetings & thanks again.

Jesper