Soundfont technology out-of-date?

[learjeff]

Gunnare, I agree that it's not dead yet! SF2 is still a useful format, and sfz is definitely a great program, especially for a free one.

Marce, I've never had APE corrupt data. However, there have been version incompatibilities from time to time, which is a bad thing. Based on my tests, APE compresses all files smaller.

Olivier Giuillion has interesting points, but he's being unnecessarily negative. The problems he mentions can easily be overcome. First of all, he's only talking about "internal" formats, and I was discussing external formats, with the possiblity of using it internally (player software's option). Random access is an issue, but only within "windows" (much like FFT processing) at any given time interval, and I feel that relatively efficient code could be written to handle it. I suspect the buffer management code would be relatively insignificant compared to the pitch shifting algorithms.

Second, there are ways to handle loop points. One way is to save the loop as a separately encoded section so that it's easy to start from.

Finally, Olivier's claim that you only get a 30% increase in available memory is terribly pessimistic. APE is very efficient on relatively simple waves (typical of single-note samples, even for relatively complex sounding instruments like piano or violin). I expect a reduction to 40%, or as low as 25%. That corresponds to a 150% to 300% increase available sample memory. For single-voice samples, the 25% value is typical. So, he's discounting an idea that could give you almost 4 times as much sample space.

[learjeff]

Back to mid-side encoding (also called "joint stereo", btw). I think there's another advantage that I neglected above. Previously, I figured that the only real advantage to mid-side is the reduced memory and CPU usage for a mono preset in a file that supports both mono and stereo.

With mid-side encoding, I could much more easily find loop points while retaining the excellent sum-to-mono characteristics. I'm currently working on a looped, stereo version of my Rhodes soundfont. The full-length stereo version sounds almost exactly what I wanted (better in many ways than I've gotten any other way, including a real Rhodes and a Roland Chorus Echo tape loop machine). The stereo image is built using pitch-shift doubling. The effect is applied with mid-side technique, meaning it's added to left and subtracted from the right. Therefore, when summed to mono, I get the original, unaffected sound -- exactly what I want in that respect.

Unfortunately, it's 160MB. Trying to loop it is giving me hives. It seems I have three choices:

1) long samples yield nice image and sum-to-mono but use a lot of memory
2) very short loops yield good sum-to-mono characteristics, but the image stops moving once it hits the loop
3) longer loops yield nicer, moving images, but when summed to mono, the sound warbles.

If sfz format was expanded to include support for mid-side-encoded wave files, I'd probably abandon sf2 format and switch to sfz format -- at least, for stereo fonts.

[Vonbrucken]