Requantise to 16-bit
Drop an audio file and export a WAV whose samples sit on the 16-bit integer grid and whose header declares that width. Triangular dither is applied before rounding.
Why rounding alone is worse than noise
Cutting a 24-bit or float sample down to 16 bits means picking the nearest available value. Done plainly, the error left behind tracks the signal that produced it, and an error correlated with the signal is heard as distortion rather than as hiss.
It is worst exactly where it is least welcome. A fade to silence, a reverb tail, or a quiet passage puts the signal down near the bottom of the grid, where the steps are large relative to what is left.
Adding a small amount of noise before rounding breaks that correlation. The error becomes a steady, signal-independent hiss, which the ear ignores far more readily than it ignores distortion that moves with the music.
Triangular rather than rectangular
Rectangular dither decorrelates the mean of the error but leaves its variance moving with the signal, which is audible as noise that breathes in and out. Triangular dither removes the modulation completely.
It costs about 4.8 dB more noise floor than the undithered case. Against a theoretical -98 dB floor at 16 bits, that is a trade worth making every time.
Enough range for delivery
16-bit provides more dynamic range than any real playback environment can reproduce. Room noise, amplifier noise, and the listener’s own hearing all sit well above the floor.
Higher depths earn their place during production, where each processing stage rounds again and the errors accumulate. At the end of that chain, one careful conversion to 16-bit is all that is needed.
Before you convert
Dither once, at the very end. If the file will be processed, resampled, or re-encoded afterwards, the dither noise is reshaped by whatever comes next and the care taken here is wasted. Convert last.
Other depths
24-bit suits masters that will be processed again. 8-bit is for deliberate lo-fi. To choose the dither type or add noise shaping, use the full Bit Depth Converter.