Part 1 of (maybe) 3 Some time ago, a question was asked on LAU regarding resampling of an audio library. I strongly recommend the following, assuming that you decide to resample (and that's your call, not anyone else's): If you can at all avoid it, *don't* use sinc resamplers (for example, sndfile-resample and resample-1.7) for bulk rate conversions at constant rates, especially for resampling something of great value such as an audio library which you plan to use over and over again on future projects. To be most certain you've got a good library at the new sample rate, re-record the original sources or purchase a competently produced new library. ------------------------ In answering this question about resampling an entire audio library, Steve Harris (prolific author of LADSPA plugins for which we are all grateful) posted: "I would say that a high quality sinc resampler will be the best quality." Upon contacting Steve and exchanging several emails, it became clear that Steve was actually talking about the theoretically ideal case, not about the algorithm developed and described by Julius O. Smith III of CCRMA and implemented in such codes as sndfile-resample and resample-1.7. Steve says he doesn't have enough experience with these programs to either agree or disagree with claims made about them. For many readers, "sinc resamplers" probably does refer to these programs, so I think it's important to attempt once again to put them into perspective: Julius O. Smith III of CCRMA, in his PDF document describing a sinc- function based algorithm in detail, himself mentioned the suitability of this particular algorithm for situations wherein the sampling rate was *changing* and *arbitrary* such as audio scrubbing, a real-time application of resampling. Resampling an audio library is an application of *bulk conversion* to a *fixed sample rate* and is very different, therefore, as is easily arguable, an unsuitable application for Smith's algorithm, therefore unsuitable for sndfile-resample and resample-1.7. I may demonstrate some problems with bulk conversions later such as I did for Steve, but first should point out that anyone who is familiar with Fourier analysis and synthesis should appreciate immediately why this is the case upon reading Smith's description of the algorithm: As described in [1] it utilizes very small, comparatively infinitesimal windows (compared to the infinite signals of the Sampling Theorem) and is of order n squared. For someone with the proper education and/or experience, this is really *all that is necessary to know* to predict problems with speed or accuracy --- or both of these. ------------------------ Reference [1] Julius Orion Smith III. Digital Audio Resampling Home Page. http://www-ccrma.stanford.edu/~jos/resample/. This reference cites (after reordering author/title): J. O. Smith and P. Gossett. Flexible Sampling-Rate Conversion Method. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing, March 1984 (ICASSP-84), Volume II, pp. 19.4.1-19.4.2, IEEE Press. Note the word "flexible."
