Dithering and Noise Shaping in Digital Audio: Eliminating Quantization Distortion
Understand digital audio resolution: bit depth, quantization error, why adding random noise eliminates harmonic distortion, and psychoacoustic noise shaping curves.
Safety First: High gain settings on dither listening checks should be monitored carefully to prevent sudden loud passages when re-enabling musical audio.
In digital audio engineering, dithering is one of the most counter-intuitive concepts: in order to make digital audio sound cleaner and more transparent, engineers intentionally inject a precise layer of random analog-style noise into the signal.
When audio is converted from a high bit depth (such as a 32-bit floating point DAW mixbus or a 24-bit studio master) down to 16-bit consumer delivery (Red Book CD or standard streaming formats), failure to dither introduces harsh, correlated quantization distortion.
1. The Anatomy of Quantization Error
Digital audio represents sound waves by slicing continuous analog voltage into discrete numerical steps. The bit depth dictates the total number of discrete voltage levels available:
- 16-bit: $2^{16} = 65,536$ voltage levels (96.3 dB theoretical dynamic range).
- 24-bit: $2^{24} = 16,777,216$ voltage levels (144.5 dB dynamic range).
When reducing 24-bit audio to 16-bit, each sample’s bottom 8 bits must be discarded. If you simply round or truncate the numbers, any subtle acoustic signal whose amplitude falls below 1 Least Significant Bit (LSB) cannot be accurately mapped:
$$\text{Error}(n) = x(n) - Q(x(n))$$
Because this rounding error is directly correlated with the input signal itself, it does not sound like gentle white background hiss. Instead, it generates correlated distortion harmonics and harsh digital “granulation noise.” During quiet musical fades or reverberation tails, truncation causes the audio to abruptly sputter, buzz, and gate into silence like an ugly digital bit-crusher.
2. What Dither Does: Linearizing the System
Dither is a mathematically controlled, microscopic noise source injected into the signal immediately prior to rounding.
In their seminal 1984 JAES paper, John Vanderkooy and Stanley Lipshitz proved that applying the correct dither waveform completely decorrelates the quantization error from the audio signal:
- The harsh, non-linear harmonic distortion is completely eliminated.
- The distortion is converted into a constant, benign, stationary background hiss.
- Crucially, dither allows signals far below the digital noise floor to remain perfectly audible and uncorrupted.
With proper dither applied to a 16-bit file, a human ear can easily resolve a pure sine tone playing at -115 dBFS — nearly 20 dB below the theoretical 96 dB 16-bit noise floor! The dither noise acts as an acoustic carrier, modulating the probability of the LSB flipping between 0 and 1.
3. Dither Probability Densities: RPDF vs. TPDF
Not all noise is equal for dithering. Digital signal processors use specific probability density functions (PDF):
Rectangular Probability Density Function (RPDF)
- Generated with a single random number generator spanning 1 LSB.
- Flaw: While it eliminates harmonic distortion of the audio signal, the noise floor itself continues to modulate in volume depending on signal amplitude (noise modulation).
Triangular Probability Density Function (TPDF)
- Generated by summing two independent uniform random noise sources spanning 1 LSB each.
- The Audio Standard: TPDF dither produces a triangular probability distribution between -1 and +1 LSB. It guarantees zero noise modulation and zero harmonic distortion. The noise floor remains utterly silent, stationary, and sonically invisible.
4. Psychoacoustic Noise Shaping
While TPDF dither solves the distortion problem, it raises the broadband noise floor of a 16-bit file by approximately 4.8 dB.
Noise shaping uses a negative feedback loop to sculpt the frequency distribution of the dither noise, pushing acoustic noise energy away from the frequencies where human hearing is most sensitive (2 kHz to 5 kHz) and pushing it up into the near-inaudible ultrasonic spectrum (14 kHz to 20 kHz):
- Fletcher-Munson Psychoacoustic Curve: Because human hearing sensitivity drops by over 30 dB at 18 kHz compared to 3 kHz, hiding dither noise in the 15–20 kHz band renders it virtually inaudible.
- Perceptual Dynamic Range: Advanced psychoacoustic noise shaping curves (such as POW-r 3, Sony Super Bit Mapping, or Apogee UV22HR) achieve an effective perceptual dynamic range exceeding 115 dB from a standard 16-bit digital file.
When to apply dither:
- Rule 1: Only dither when reducing bit depth (e.g., 32-bit float to 24-bit, or 24-bit to 16-bit).
- Rule 2: Dither should be the very last process on your master export chain.
- Rule 3: Never dither twice; dithering already-dithered 16-bit files simply piles unnecessary noise onto the recording.