Bit depth and quantization noise
Bit depth determines how many levels each sample can take: 16 bit = 65,536 levels, theoretical dynamic range of about 96 dB; 24 bit = about 16.7 million levels, about 144 dB theoretical; 32-bit float = floating point, practically impossible to overload inside the DAW. When you reduce the bit depth (for example from 24 or 32 float to 16 bit for CD), the values are rounded: the rounding error is quantization noise, which in quiet passages (reverb tails, fade-outs) becomes signal-correlated distortion (a grainy “scratch”).
What dither is
Dither is very low-level noise added before the bit reduction. It makes the quantization error random and uncorrelated with the signal: instead of distortion you get a very low constant hiss (around −90/−96 dBFS at 16 bit), and details below the last bit remain audible “through” the noise.
- TPDF (triangular probability density function): the standard dither, transparent and free of modulation.
- Noise shaping: moves the noise energy toward the frequencies where the ear is least sensitive (above 15 kHz), reducing the perceived noise by several dB. Commercial examples: POW-r (types 1, 2, 3), iZotope MBIT+, the dithers in FabFilter Pro-L 2 and Ozone.
- More noise shaping is not always better: with many conversions or further processing, the noise pushed upward can accumulate.
Practical rules
- Dither only once, as the last step, when reducing bit depth (typically 24 → 16 for CD).
- No dither if you export to 32-bit float.
- When exporting to 24 bit from a float project, dither is theoretically correct but its effect is at levels below the noise floor of any playback system; many engineers apply it anyway, others don't: neither is a mistake.
- After dither, do not process any further (no gain, EQ, or normalization), otherwise it must be redone.
- For streaming you usually deliver at 24 bit: the platform will do its own conversions.
- Fade-outs and tails: dither matters most right there.
Sample rate conversion (SRC)
Going from 48 kHz (or 96 kHz) to 44.1 kHz requires an anti-aliasing filter and resampling. A good SRC is transparent; a poor one introduces aliasing or loss of highs. Tips:
- Work at the sample rate of the main delivery when possible (44.1 or 48 kHz for music; 48 kHz for video).
- If you must convert, use high-quality algorithms (iZotope RX, SoX, the DAW's SRC at maximum quality) and convert before the final limiter, then check the true peak again (conversion can alter inter-sample peaks).
- Typical end-of-chain order: processing → SRC → limiter (true peak) → dither → export.
In FL Studio
In the export window choose WAV with 16, 24, or 32-bit float depth; for 16-bit export enable the dithering option (FL offers dither with noise shaping). The resampling quality (interpolation) is set in the export options (choose the highest quality for the final render). Many mastering limiters (Pro-L 2, Ozone Maximizer) have built-in dither: in that case turn off the export one so it isn't applied twice.
Example
Master of an EP in A minor: project at 48 kHz/32 float, delivery for streaming at 44.1 kHz/24 bit and for CD at 44.1/16. Chain: EQ → compression → SRC 48 → 44.1 (high quality) → limiter with true peak −1 dBTP → 24-bit export (light TPDF dither) for streaming; from the same chain, 16-bit export with noise-shaped dither for the CD DDP. Final check: the final fade-out listened to on headphones at high gain to verify there is no grainy distortion.
Exercise
Export a long fade-out at 16 bit without dither and with TPDF dither; raise both by 40 dB (for listening only) and compare: can you hear the difference between distortion and hiss?
Links
“Formats and Delivery”, “Limiting, Clipping, and True Peak”, “Mastering for Streaming”, “Gain Staging”.