noise-shaped (Wannamaker 3). TPDF dither with a 3-tap error-feedback filter that pushes the noise energy into the high frequencies (>14 kHz) where hearing is less sensitive. perceived noise is lower than TPDF.
[why dithering matters]
truncation only
noise-shaped dither
a -80 dBFS sine wave at 1 kHz reduced from float to 16-bit. when you reduce 24-bit (or 32-bit float) to 16-bit, the bottom 8 bits of each sample are discarded. without dither this creates a stepped quantisation distortion most audible on quiet sounds and fade-outs (the red wave on the left). dither adds a tiny amount of noise before truncating, which converts the harsh distortion into a much less objectionable, broadband noise floor.
[dither noise spectrum]
dither noise floor for each type. TPDF (cyan) is flat — the canonical -93 dBFS noise floor. noise-shaped (amber) starts lower in the audible range and rises sharply above 14 kHz. flat (purple) is similar to TPDF but slightly less effective at masking distortion. truncation (red) is shown for reference — the “noise” here is actually quantisation distortion correlated with the input.
[apply to your audio]
drop a 24-bit (or 32-bit float) audio file
reduces to 16-bit with the selected dither · capped at 180s