Talk:Dither

Q: Perhaps it would be better to discuss what quantization errors are before observing that dither is intentional noise.
 * Gale 05Jan11: Maybe, but there is a link to Wikipedia for quantisation errors and it is explained as "rounding errors". I think it may be better to keep the first sentence simple so an inexperienced user at least gets the general idea therefrom.

Q: You certainly do need to dither down to 24 bits when going from 32-bit float to 24-bit PCM wave! You don't need dither for the inverse conversion (24-bit PCM to 32-bit float). Simply discarding the 8 exponent bits of the 32-bit float is not good (or typical) -- I hope this is just an error in this wiki and not what audacity is really doing.
 * Gale 05Jan11: Steve Daulton from the Forum wrote that text so I've let him know about your comment, however definitely, no dither is applied when Audacity downsamples from 32-bit float to 24-bit PCM.
 * Steve Daulton 12Jan11: I'm not sure that everything attributed to me is what I actually wrote, however I'm looking into this and finding some interesting results. The suggestion of "simply discarding the 8 exponent bits of the 32-bit float" is certainly incorrect. Signed integer uses twos compliment for negative numbers and float uses sign and magnitude. I've also discovered a bug in Audacity 1.3.12 where the highest possible positive sample value for 24-bit sample is 7F FF FE when it should be 7F FF FF. This appears to have been corrected in the current Audacity 1.3.13 version, though I've only tested this on Linux.

In Audacity 1.3.13 on Linux, dither is applied to 24-bit data when going from 32-bit float to signed 24-bit if enabled in preferences (though at a much lower level than when going from 32-bit float to signed 16-bit).


 * Steve Daulton 13Feb11:I've attempted to correct the details regarding 24-bit conversion. When converting the sample format of a track within Audacity, dither is applied in 3 cases. 32-bit -> 24-bit, 32-bit -> 16-bit and 24-bit -> 16-bit. When exporting data I believe what is intended to happen is that the audio data is rendered at 32-bit, then converted to the sample format required by the file format, thus if the file format is 16 or 24-bit the data will have dither applied (assuming that dither is enabled in Preferences). If dither is disabled in Preferences, then dither should never be applied.

In Audacity 1.3.12 there are inaccuracies when converting from some 24-bit integer values to 32-bit float. This appears to only apply to some values over about +/-0.8 (linear scale), but it does mean that some statements in the wiki page are not entirely true for the current version of Audacity.