/ˈdɪð.ər.ɪŋ/
Dithering is the process of adding a precise, low-level noise signal to audio before reducing its bit depth, replacing harsh quantization distortion with a softer, more musical noise floor and preserving fine detail in quiet passages.
Every master you've ever delivered lives or dies in a rounding error — dithering is the one tool that turns that rounding error into something a human ear will never complain about.
Dithering is the intentional addition of a low-amplitude, statistically random (or shaped) noise signal to a digital audio file immediately before reducing its bit depth. When a 24-bit or 32-bit floating-point session is exported at 16-bit resolution — the standard for CD, streaming-optimized delivery, and legacy distribution — each audio sample must be rounded to the nearest available value in the coarser word length. Without dithering, that rounding produces quantization distortion: a signal-correlated, harmonically offensive artifact that becomes audible as a harsh, grainy texture during reverb tails, fade-outs, and quiet ambient passages. Dithering replaces this correlated distortion with a consistent, signal-independent noise floor that the ear tolerates far more graciously than the alternative.
The fundamental mechanism exploits a psychoacoustic and mathematical principle: randomizing rounding errors breaks the correlation between the distortion and the audio signal. Without dither, quantization error follows the waveform precisely, which means its harmonic structure is directly related to the music — a relationship the auditory cortex is exquisitely sensitive to. With dither, the error becomes statistically white or shaped noise, which the ear perceives as simple tape hiss rather than digital grit. The perceptual difference is dramatic: a 16-bit signal without dither can exhibit quantization artifacts equivalent to roughly −96 dBFS dynamic range, but with harsh, structured character; a properly dithered 16-bit signal exhibits approximately the same noise floor but with a smooth, benign quality that is effectively inaudible in most listening contexts.
It is critical to understand where dithering belongs in the signal chain. Dithering is applied exactly once, as the very last processing step before bit-depth reduction. It must never be applied mid-chain, between two processes that both operate at high internal precision — doing so accumulates noise with no benefit. In practice, this means dithering is applied at the mastering stage output, after all EQ, compression, limiting, and level adjustments are complete. If you are bouncing a 24-bit mix to send to a mastering engineer, dithering is unnecessary and should be disabled; the mastering engineer will apply dither at the final export. If you are self-mastering and delivering a 16-bit file, dithering is mandatory.
Modern DAWs universally operate internally at 32-bit or 64-bit floating-point precision, which means quantization is not a meaningful concern within the session itself — floating-point arithmetic provides sufficient resolution that dither adds no value during internal processing. The critical moment arrives only at the final render, when floating-point audio must be committed to an integer word length for delivery. At this moment, a properly selected dithering algorithm — ideally with noise shaping calibrated to the target medium — is the difference between a transparent, professional master and one with artifacts that reveal themselves under critical listening conditions, particularly on high-quality headphones and audiophile playback systems.
Despite being a mathematically small intervention — dither noise typically sits between −93 and −120 dBFS depending on the noise-shaping curve — its perceptual impact is significant enough that it has been a standard part of professional mastering workflow since the CD era. Engineers including Bob Ludwig, Bernie Grundman, and Greg Calbi apply dithering as a non-negotiable final step on every 16-bit master. Understanding dithering is not optional knowledge for a serious producer: it is a foundational competency that separates work that sounds professional at every playback level from work that degrades subtly but unmistakably in quiet passages.
At the mathematical core, dithering works by modifying the quantization process. In standard truncation or rounding without dither, each sample value in the high-bit-depth signal is mapped to the nearest representable value in the target bit depth. For a 16-bit integer, there are 65,536 discrete steps between the minimum and maximum level; for a 24-bit signal, 16,777,216. When a sample falls between two 16-bit steps, it is rounded to the closest one — a deterministic operation that creates an error signal with a fixed relationship to the input. Because this error is signal-correlated, it manifests as harmonic distortion and intermodulation artifacts that are perceptible even at very low levels, especially on sustained tones and decaying reverb tails where the signal amplitude approaches the quantization step size.
Dithering adds a noise signal to the audio before quantization occurs. This noise randomizes the rounding decision: instead of always rounding to the nearest step, the combination of audio and dither noise causes the quantizer to oscillate between neighboring steps in a statistically random pattern. Over many samples, the average value converges on the true signal, effectively encoding amplitude information below the theoretical noise floor of the target bit depth. This is the same statistical principle used in sigma-delta modulation and pulse-width modulation — rapid dithering between two states that averages to an intermediate value. The result is that fine detail in reverb tails, room ambience, and quiet musical passages is preserved as a gentle noise modulation rather than lost to harsh truncation.
Noise shaping extends the basic dithering concept by spectrally sculpting the added noise so that it is concentrated in frequency regions where human hearing is least sensitive. The equal-loudness contours described by Fletcher and Munson — and later refined into the ISO 226 standard — reveal that the ear is most sensitive between approximately 2 kHz and 5 kHz, and significantly less sensitive below 200 Hz and above 15 kHz. Noise-shaping algorithms exploit this by using an error-feedback loop: the quantization error from each sample is filtered and fed back into the next sample's dither calculation, effectively pushing noise energy into the less-sensitive high-frequency region. The perceptually weighted noise floor can drop 10–20 dB below simple flat dither, meaning the audible noise is dramatically quieter even though the total noise energy may be similar or even slightly higher.
The three dominant noise-shaping standards in professional mastering are Type 1 (flat dither, no shaping), Type 2 (moderate high-frequency emphasis, suitable for general use), and the POW-r (Psychoacoustically Optimized Wordlength Reduction) algorithm suite developed by the POW-r Consortium — a group that included engineers from Waves, Weiss, and other mastering hardware manufacturers. POW-r Type 1 applies no shaping; Type 2 applies moderate shaping optimized for 44.1 kHz; Type 3 applies aggressive high-frequency noise shaping that yields the lowest audible noise floor at 44.1 kHz sample rates. iZotope's MBIT+ algorithm and Apogee's UV22HR are additional proprietary implementations that use psychoacoustic models to optimize noise placement differently. The choice between algorithms affects primarily the audible noise character, not the signal accuracy, and in most real-world masters — where program material rarely drops below −60 dBFS — the differences are subtle enough to require trained critical listening to discern.
Practically, the signal chain for correct dithering looks like this: audio is processed entirely within the DAW's high-precision internal engine (32- or 64-bit float), all processing is completed including final limiting, the output ceiling is set (typically −0.1 to −0.3 dBFS true peak for streaming), dither is inserted as the last plugin or export option, and the file is rendered to 16-bit integer. The dither noise is encoded permanently into the file at that point. Any subsequent processing of the 16-bit file — such as sample-rate conversion, encoding to MP3, or re-importing into a new session — should not involve re-dithering unless another bit-depth reduction is taking place, as stacking dither passes accumulates noise unnecessarily.
Diagram — Dithering: Signal flow diagram showing 32-bit float audio path through dither and noise shaping to 16-bit integer output, with quantization error comparison.
Every dithering — hardware or plugin — operates on the same core parameters. Know these and you can work with any implementation.
Common options include flat (TPDF — Triangular Probability Density Function), POW-r Type 1/2/3, UV22HR, and MBIT+. Each algorithm places noise energy differently across the frequency spectrum. POW-r Type 3 and MBIT+ offer the lowest perceptible noise floors at 44.1 kHz by aggressively concentrating noise above 15 kHz, making them the preferred choices for music masters.
Noise shaping curves range from flat (no shaping, Type 1) to aggressive high-frequency emphasis (Type 3, moderate, steep). Aggressive shaping yields a perceptually quieter floor — sometimes 10–20 dB quieter in the 2–5 kHz critical hearing band — but increases total high-frequency noise energy. At 44.1 kHz, aggressive shaping is safe; at 48 kHz or higher, moderate shaping is preferred because the Nyquist frequency is higher and noise concentration is less effective.
The most common target is 16-bit (CD, most streaming) with a theoretical noise floor of −96 dBFS. Some workflows target 24-bit (hi-res streaming via Tidal, Apple Music Lossless, Qobuz), where the noise floor is −144 dBFS and dithering noise is so far below audibility that many engineers omit it. Delivering a 24-bit file without dither is generally acceptable; delivering a 16-bit file without dither is a technical error.
Standard TPDF dither uses a peak amplitude equal to ±1 least-significant bit (LSB) of the target word length. For 16-bit audio, this is approximately −96 dBFS. Some noise-shaped algorithms effectively apply noise at sub-LSB amplitude in the audible band while concentrating energy above 15 kHz. Increasing dither amplitude beyond one LSB provides no additional benefit and raises the audible noise floor unnecessarily.
Dithering must be placed after all other processing — the absolute final step before bit-depth reduction. Inserting dither before a limiter, for example, would cause the limiter to process the dither noise and alter its statistical properties, defeating its purpose. In most DAWs, dither is either the last plugin on the master bus or selected in the export/bounce dialog, ensuring it is applied at the final conversion stage.
At 44.1 kHz, the Nyquist ceiling is 22.05 kHz, giving noise-shaping algorithms a narrow band above 15 kHz to concentrate energy. Aggressive curves (POW-r 3) exploit this effectively. At 48 kHz, the extra bandwidth allows slightly gentler curves while achieving similar perceptual results. At 88.2 or 96 kHz, dithering to 24-bit is common and noise shaping provides minimal additional benefit given the already-low perceptual noise floor.
Session-ready starting points. These settings apply to the final 16-bit export step; always disable dither when sending stems or mixes to a mastering engineer.
| Parameter | General | Drums | Vocals | Bass / Keys | Bus / Master |
|---|---|---|---|---|---|
| Dither algorithm | POW-r 3 or MBIT+ | POW-r 3 | POW-r 3 or MBIT+ | POW-r 2 or MBIT+ | POW-r 3 or UV22HR |
| Target bit depth | 16-bit | 16-bit | 16-bit | 16-bit | 16-bit (CD/streaming) |
| Noise shaping | Moderate–Aggressive | Aggressive (Type 3) | Aggressive (Type 3) | Moderate (Type 2) | Aggressive (Type 3) |
| Apply dither? | Yes — final export only | Yes — final export only | Yes — final export only | Yes — final export only | Yes — once, last step |
| Dither for 24-bit deliverable | Optional / omit | Omit | Omit | Omit | Optional |
| Mid-chain dithering | Never | Never | Never | Never | Never |
| Dither before mastering engineer | No — send 24-bit | No — send 24-bit | No — send 24-bit | No — send 24-bit | No — send 24-bit |
These settings apply to the final 16-bit export step; always disable dither when sending stems or mixes to a mastering engineer.
The mathematical foundations of dithering predate digital audio by several decades. During World War II, engineers working on early analog-to-digital computing systems — including work at MIT's Whirlwind project — observed that mechanical and electrical computing systems produced more accurate average outputs when subjected to controlled vibration or noise. The term "dither" itself derives from Old English and Middle English roots suggesting trembling or indecision, and was adopted in engineering to describe the deliberate introduction of noise to linearize quantization systems. By the 1950s, the concept was documented in control systems literature, and by the 1960s it appeared in early pulse-code modulation (PCM) research at Bell Laboratories, where engineers were developing the theoretical framework for digital audio transmission.
Practical dithering for consumer digital audio emerged with the introduction of the Compact Disc in 1982. The Sony PCM-1600 and PCM-1610 professional digital audio processors, which were used to master many of the first CD releases, operated at 16-bit resolution, and Sony's engineers including Toshitada Doi recognized early that unmodified quantization at 16-bit depth produced audible artifacts in quiet passages. Thomas Stockham at the University of Utah had also documented quantization distortion in early digital recording experiments in the 1970s. The first commercial recordings that systematically applied dithering at mastering included productions from Sony's early CD catalog, though specific documentation of which titles used which algorithms is sparse, as the practice was not yet standardized or widely discussed publicly.
The 1990s saw the formalization of noise shaping as an extension of basic dithering. Robert Wannamaker, Stanley Lipshitz, and John Vanderkooy at the University of Waterloo produced a landmark series of academic papers between 1984 and 1993 that rigorously characterized optimal dither and noise-shaping theory, including proof that Triangular Probability Density Function (TPDF) dither completely decorrelates quantization error from the input signal. These papers became foundational reading for mastering engineers and audio DSP developers. Concurrently, Apogee Electronics developed UV22, their proprietary ultrasonic noise-shaping dithering algorithm, which was used on thousands of commercial releases throughout the 1990s and became particularly associated with the work of mastering engineers at the Mastering Lab in Los Angeles. UV22HR, the updated version, remains in use today.
The POW-r Consortium — a collaboration that included Waves Audio, Weiss Engineering, Prism Sound, and other mastering hardware manufacturers — codified and licensed the POW-r (Psychoacoustically Optimized Wordlength Reduction) algorithm suite in the late 1990s. POW-r's three algorithm types, developed with input from mastering engineers and psychoacousticians, became the industry standard implemented in Waves plug-ins, hardware mastering converters from Weiss and Prism, and licensed to numerous DAW manufacturers. iZotope introduced MBIT+ with their Ozone mastering suite in the early 2000s, offering an additional psychoacoustically modeled option that became widely adopted by independent producers as the Ozone platform grew in popularity. By the mid-2010s, with streaming platforms standardizing on 16-bit FLAC for lossy-lossless delivery and 24-bit for hi-res tiers, the conversation around dithering expanded to include whether 24-bit deliverables require dither at all — a question now generally answered in the negative for practical purposes, given that the −144 dBFS noise floor of undithered 24-bit audio is below the threshold of any known playback system's noise floor.
For self-mastering producers, dithering is applied once, at the final bounce. The workflow is consistent regardless of genre: complete all EQ, compression, saturation, and limiting on the master bus; set the output ceiling (typically −0.1 dBFS for CD or −1.0 dBFS true peak for streaming-optimized files); insert the dithering plug-in as the absolute last processor on the master bus; then export. In Ozone or a standalone dither plug-in, select POW-r Type 3 or MBIT+ for 44.1 kHz 16-bit masters. Disable dither when bouncing 24-bit files intended for a mastering engineer. This single workflow decision — remembering to apply dither at export and remembering to disable it otherwise — accounts for a significant proportion of technical errors in independent releases.
When mastering for multiple formats simultaneously — a common requirement for artists delivering to CD duplication houses, streaming platforms, and hi-res download stores in the same session — the correct approach is to create separate exports for each format. The 24-bit / 96 kHz hi-res version requires no dithering. The 16-bit / 44.1 kHz CD master requires dithering. Some mastering engineers prefer to derive the 16-bit version from the 24-bit render (with sample-rate conversion performed first) rather than from the original session, ensuring the two formats are bit-for-bit equivalent except for the final dithering step. Plug-ins like iZotope Ozone's Dithering module and the Waves L2 Ultramaximizer with its built-in IDR (Increased Digital Resolution) dithering support this workflow directly within the master bus chain.
Producers working in electronic music, ambient, and classical — genres where reverb tails, fade-outs, and quiet passages are common — benefit most from careful dither algorithm selection. In a quiet electronica track where a pad decays over eight seconds into silence, undithered truncation can produce a characteristic "grainy" texture as the signal level drops below approximately −80 dBFS. With aggressive noise shaping enabled, the same decay dissolves into smooth, barely perceptible hiss that most listeners will attribute to tape-like warmth rather than digital artifact. Hip-hop, EDM, and heavily compressed pop masters with noise floors well above −60 dBFS are less sensitive to dither algorithm choice, though proper dithering remains correct practice.
A common advanced workflow involves using a hardware mastering chain — Manley Vari-Mu, API 2500, Dangerous Master — feeding into a high-quality analog-to-digital converter such as the Prism Sound Orpheus, Lynx Aurora(n), or Apogee Symphony. These converters perform the analog-to-digital conversion at 24-bit depth, and their internal dithering or noise-shaping circuits handle the conversion to the converter's native word length. When the resulting 24-bit file is then rendered to 16-bit in the DAW, a second dithering pass is applied in software. The converter's internal dithering and the software dithering are separate operations at separate bit-depth boundaries — the converter dithers from its internal high-precision representation to 24-bit, and the software dithers from 24-bit to 16-bit — and both are correct and necessary steps in the chain.
One email a week. The techniques behind the terms — curated by working producers, not algorithms.
Abstract knowledge becomes practical when you can hear it in music you know. These tracks demonstrate dithering used intentionally, at specific moments, for specific purposes.
This recording is frequently cited in mastering education as an example of pristine 16-bit dithering. The opening guitar introduction begins at approximately −50 dBFS, well within the range where quantization distortion would be audible without proper dither. Listen on high-quality headphones as the guitar decay trails off near the end of each phrase — the noise floor is smooth and continuous, with no granular texture or harmonic artifacts that would indicate undithered truncation. Greg Calbi's use of UV22HR or POW-r-class dithering at Sterling Sound is evident in how cleanly the quiet passages resolve.
The extended ambient decay at the end of this track descends well below −70 dBFS over approximately 20 seconds. The 16-bit CD master demonstrates effective dithering in the way the residual texture of the fading synth pads resolves — the noise character is consistent and broadband rather than signal-correlated or granular. This is a useful A/B reference: import the CD rip into your DAW, zoom into the final decay waveform at high magnification, and compare its noise behavior to a test export of your own material. The absence of quantization artifacts in the Boards of Canada fade is textbook.
The ECM CD remaster of this solo piano recording is a standard mastering benchmark. The near-silence between notes and phrases — piano is among the most demanding instruments for 16-bit resolution because of its wide dynamic range and long, low-level decay tails — reveals the quality of dithering in any 16-bit transfer. On the remaster, the noise floor during the decay of individual bass notes is smooth and unstructured. Comparing this master to amateur piano recordings truncated without dither reveals the classic quantization artifact: a buzzy, pitch-correlated degradation of the quiet tail that sounds almost like digital compression pumping.
The a cappella vocal opening of this track descends into whispered near-silence before the bass entry at 0:32. The 16-bit CD single demonstrates effective dithering in the handling of Elizabeth Fraser's breathy vocal texture at levels approaching the noise floor. Producers working with intimate vocal recordings should use this as a dither reference: the goal is that quiet sibilance and breath noise resolve as smooth white noise rather than as a structured, distorted approximation of the original signal.
Triangular Probability Density Function dither adds two rectangular noise distributions, producing a triangle-shaped probability curve that fully decorrelates quantization error from the input signal. It is the mathematically proven minimum dither type needed to eliminate signal correlation. The resulting noise is broadband and flat — essentially white noise at −96 dBFS for 16-bit — which is audible in very quiet passages. It is the safest choice for any sample rate and the correct default when no psychoacoustic optimization is required, such as for archival purposes or when delivering to a format where further processing will occur.
POW-r (Psychoacoustically Optimized Wordlength Reduction) algorithms apply an error-feedback loop to spectrally shape the dither noise, pushing energy toward the upper frequency range where human hearing is less sensitive. Type 2 applies moderate shaping suitable for 48 kHz material; Type 3 applies aggressive shaping optimized for 44.1 kHz, achieving a perceptual noise floor up to 20 dB lower than flat dither in the 2–5 kHz critical hearing band. POW-r Type 3 is the most widely recommended choice for 16-bit CD and streaming masters and is available in Waves, Ozone, and Logic Pro's bounce dialog.
Apogee's UV22HR algorithm encodes dither information at 22 kHz — near the Nyquist limit of 44.1 kHz audio — creating an ultrasonic carrier that psychoacoustically masks the noise in normal listening. UV22HR became the signature dither of the 1990s mastering community and is associated with major releases mastered at Gateway Mastering, Sterling Sound, and the Mastering Lab. Its character is often described as slightly warmer than POW-r Type 3, with less perceived high-frequency hash. UV22HR is available as a native plug-in within Apple Logic Pro and as a standalone Apogee utility.
iZotope's MBIT+ (Maximized Bit-depth) algorithm is the dithering engine built into Ozone and available as a standalone module. It applies a psychoacoustically modeled noise-shaping curve with adjustable aggressiveness (low, medium, high) and includes a setting for 20-bit output for specialized mastering scenarios. MBIT+ at its highest noise-shaping setting is generally considered competitive with POW-r Type 3 and is the default recommendation for producers working within the Ozone ecosystem. It is available in the free version of Ozone Elements, making it accessible to producers without a full iZotope subscription.
Rectangular dither adds a uniform-distribution noise signal that reduces quantization error but does not fully decorrelate it — it is statistically less optimal than TPDF but computationally cheaper. Some legacy CD mastering systems used rectangular or even no dithering, and some low-cost export dialogs still default to this mode. It is audibly inferior to TPDF or noise-shaped dithering in critical listening conditions and should be avoided for professional masters. Its only appropriate use case is in educational or diagnostic contexts where the effect of sub-optimal dithering needs to be demonstrated.
These MPW articles put dithering into practice — specific techniques, real tools, and applied workflows.