NOMN: Temporal Fine Structure Enhancer
What NOMN does to audio
In our measurements a recorded flute and violin vary by 0.25 % from one cycle to the next, a singing voice by 0.74 % and a speaking voice by 2.5 %. A tape machine and a turntable add their own drift to whatever they play. A synthesized tone varies by 0.02 %, which is the floor of the measurement, and a drum machine or sampler repeats each hit exactly (waveform correlation 1.000000 between one hit and the next). A recording of a voice keeps the voice's own variation, and the digital chain adds nothing to it.
NOMN works in two layers. The drift layer moves the whole signal slowly, in the range of tape and record-player motion. The fine-timing layer adds variation of a few microseconds, at the scale auditory science calls temporal fine structure (see Temporal fine structure below). Both are generated live by models trained on the human voice: on multilingual recorded speech, and on electroglottograph recordings, which measure the opening and closing of the vocal folds at the larynx, cycle by cycle. Neither layer loops. Juiciness sets the depth of both layers. The fine-timing amount scales the fine-timing layer.
How far NOMN moves the audio
Juiciness drift layer, RMS as speed deviation in cents, RMS fine-timing layer
0.25 25 µs 0.015 % 0.26 1.0 µs
0.5 51 µs 0.029 % 0.50 1.5 µs
1.0 101 µs 0.054 % 0.93 2.5 µs
2.5 252 µs 0.134 % 2.32 5.8 µs
5.0 505 µs 0.268 % 4.63 11.4 µs
10.0 1,008 µs 0.535 % 9.23 20.3 µs
tape model 149 µs 0.068 % 1.18 0.0 µs
record player model 306 µs 0.116 % 2.00 0.0 µs
Speed deviation is the figure a wow-and-flutter meter reports. The drift layer moves slowly (90 % of its motion is below 1.7 Hz). The fine-timing layer moves faster (90 % of its motion below 15.9 Hz) and at Juiciness 1 is 40 times smaller (2.5 µs against 101 µs). Both layers scale in proportion to their controls. Level is unchanged at every setting.
NOMN and tremolo, chorus and tape emulation
NOMN differs in the source of its motion, and the difference is measurable. A tremolo or chorus is driven by an LFO. Tape and record-player emulation follow the mechanics of a transport: once per revolution, capstan rate. A humanizer uses filtered random noise. NOMN's motion comes from models trained on the voice.
We tested whether the difference shows in the audio. We put one steady synthesized tone, three minutes long, through each process, read its timing from the tone's phase, removed its size so that only the shape of the motion counted, and asked a classifier to name the process on stretches of the recording it had not been trained on. Against an LFO, our tape model, our record player model, the open-source CHOW Tape Model plugin, and Gaussian and 1/f humanizer noise, the pieces were named correctly 52 of 52 times in each comparison. An LFO, a tape machine and a record player have one kind of motion, a change of speed. NOMN has a fine-timing layer as well.
The test establishes that NOMN is a distinct process. It does not test listener preference. The measurements, method and audio examples are on this page.
Why the voice is the model
NOMN's variation comes from the body. Its models learned timing from multilingual recorded speech and from electroglottograph recordings of the vocal folds. The voice is a sound source with no rotation rate. Across 103 languages we found the timing at this scale to be much the same, which suggests it belongs to the physics of the voice more than to any one language. In our measurements NOMN's drift is aperiodic where a transport's is periodic, and NOMN has a fine-timing layer at the scale of temporal fine structure, which a transport does not have.
If timing variation matters to listeners, and the long attachment to analog transports suggests it does, then one candidate for the variation to add is the kind the voice produces, at the resolution hearing works at. In size NOMN covers the same range: at Juiciness 1 its drift is 0.054 % in speed, beside 0.068 % for our tape model. We have not yet run the listening tests that would establish a preference, and when we do we will publish them.
Emulation takes its models from the equipment of the past. We would like tools to be designed from what is known about hearing and about the bodies that make sound.
Microsecond timing and the musical JND
JND (just-noticeable-difference) thresholds for musical timing measure how far one note has to move relative to another before a listener can consciously identify the shift in a forced-choice task. They tell us when a timing difference can be labeled. They do not give the resolution at which the auditory system processes time.
The auditory system's temporal resolution is 1,000 to 5,000 times finer (10 µs against 10 to 50 ms). There are two established lines of evidence.
The binaural pathway resolves interaural time differences (ITD) of about 10 µs. Klumpp & Eady (1956, J. Acoust. Soc. Am. 28: 859-860) measured average ITD discrimination thresholds of 9 µs for band-limited noise and 11 µs for a 1000 Hz tone across ten listeners. Brughera, Dunai & Hartmann (2013, J. Acoust. Soc. Am. 133: 2839-2855) confirmed thresholds just above 10 µs at 700 to 1000 Hz. Neurons in the medial superior olive perform coincidence detection on phase-locked spikes from each ear. The largest ITD a listener normally meets, for a sound directly to one side, is 600 to 700 µs, set by the distance between the ears, and listeners resolve angular differences of about 1 degree near the midline (Mills 1958, J. Acoust. Soc. Am. 30: 237-246). The Klumpp and Eady measurement dates from 1956.
The monaural pathway encodes the sub-millisecond structure of sounds through what auditory neuroscience calls temporal fine structure (TFS): the rapid waveform oscillations within each cochlear frequency band, as distinct from the slower envelope (ENV) superimposed on them (Moore 2008, J. Assoc. Res. Otolaryngol. 9: 399-406). TFS is carried in the timing of auditory-nerve spikes that phase-lock to individual cycles of the waveform for components up to several kilohertz (Palmer & Russell 1986, Hearing Research 24: 1-15). This is the standard model of how the auditory periphery encodes time.
The auditory system uses TFS for the pitch of complex tones, for speech in fluctuating background noise, and for separating sources. Smith, Delgutte & Oxenham (2002, Nature 416: 87-90) built "chimaeric" sounds in which the envelope of one signal was combined with the TFS of another, and listeners perceived pitch and source location from the TFS. TFS coding applies to whatever the cochlea receives, including headphones and speakers playing recorded music. It runs continuously and without conscious report, which is why musical JND studies do not measure it. JND measures what listeners can report about that processing.
Our question is whether technology that generates audio for human listening should operate at the resolution of the sensory system it serves. A/B discrimination in a controlled trial is a separate question.
The audio industry has answered the first question the same way for decades. Studios record at 96 kHz or 192 kHz so the production chain introduces no artifacts near the limit of the system, whether or not listeners can distinguish those rates from 48 kHz on a given track. Mastering engineers specify word-clock jitter well below classical audibility thresholds, so the clock is never the limiting element.
NOMN follows the same principle. A synthesized source played from a digital clock has less timing variation than any acoustic source we measured: 0.02 % from cycle to cycle, against 0.25 % to 2.5 % for recorded instruments and voices. The industry already engineers sample rate, bit depth, jitter, latency, frequency response and distortion to the resolution of hearing. We apply the same practice to timing variation, where the relevant quantities also sit below the threshold of conscious labeling.
Sound localization and the separation of voices in a crowd both depend on temporal resolution far finer than the musical JND.
Temporal fine structure
The TFS framework has been developed in the auditory science literature since 2002. Moore (2008, J. Assoc. Res. Otolaryngol. 9: 399-406) reviews its role in pitch perception, masking and speech perception. Smith, Delgutte & Oxenham (2002, Nature 416: 87-90) used chimaeric sounds to show that listeners rely on TFS for pitch and localization and on ENV for speech recognition in quiet. Later work (Lorenzi et al. 2006, PNAS 103: 18866-18869; Hopkins & Moore 2009, J. Acoust. Soc. Am. 125: 442-446) showed that TFS sensitivity is critical for speech perception in noise, and that reduced TFS sensitivity is a major factor in the difficulty hearing-impaired listeners have with speech in noise.
TFS is the established vocabulary for the scale NOMN's fine-timing layer works at. NOMN operates in a well-mapped region of the auditory science literature.
Existing TFS research concentrates on loss: how hearing-impaired listeners lose TFS sensitivity, how cochlear implants fail to deliver TFS, how aging degrades TFS processing. NOMN starts from listeners whose TFS processing is intact and asks what timing structure playback technology should present to them.
Scope. NOMN has two layers. Its drift layer moves the whole signal by 25 to 1,008 µs RMS across the Juiciness range, slowly, in the range of a tape transport. Its fine-timing layer works at 1.0 to 20.3 µs across the same range. The fine-timing layer is the one at fine-structure scale. We use the term for the scale of the layer and for the auditory mechanism it addresses, and we do not publish how the layer is made.
We do not claim that digital audio is missing temporal fine structure, or that NOMN restores something the format lost. A PCM recording carries the fine structure of the in-band content of the signal. Our claim is that digital synthesis and digital playback add no timing variation of their own, and that NOMN adds a layer of variation learned from the voice. It works the same way whatever fine-structure detail the source file contains.
Latency and sub-JND timing
Jack, Mehrabi, Stockman & McPherson (2018, Music Perception 36: 109-128) tested professional percussionists and amateur musicians on a digital percussion instrument at 0 ms, 10 ms, 10 ms with ±3 ms jitter, and 20 ms of latency. Both groups rated zero latency significantly higher than the 10 ms with jitter and 20 ms conditions, and the professionals changed their timing under added latency. Schmid, Ambros, Bogon & Wimmer (2024, Proc. Audio Mostly, ACM) measured the just-noticeable difference for added audio latency across 37 listeners and found a mean of 49 ms at 0 ms base latency and 27 ms at 64 ms base latency. Asynchronies of 30 to 50 ms occur between players in ensemble performance (Rasch 1979, Acustica 43: 121-131).
An acoustic drum reaches the drummer's ears 2 to 3 ms after stick contact, a delay set by the speed of sound over the distance from drum to head.
Interface makers do not target the 27 to 49 ms at which listeners report a difference. They engineer latency well below the threshold of report, as studios do with sample rate and as clock designers do with jitter.
These studies work at 10 to 49 ms. They support the practice of engineering below the threshold of conscious report, and they say nothing about microseconds. The evidence at the microsecond scale is the interaural and fine-structure research in the two answers above.
Speaker and room effects
The difference is in structure. Room and speaker convolution is content-blind and stationary: the impulse response is fixed for a given listening position, and the reverb tail of a snare hit and of a sustained vocal note get the same treatment.
The auditory system has well-documented machinery for separating direct sound from reflections. Wallach, Newman & Rosenzweig (1949, American Journal of Psychology 62: 315-336) described the precedence effect: when two identical sounds arrive within a few milliseconds of each other, the listener hears one fused sound at the position of the first wavefront, and the later reflections contribute little to perceived location. Auditory scene analysis (Bregman 1990, MIT Press) extends this to the grouping cues by which hearing organizes sound into sources distinct from their environment. Later reviews (Litovsky et al. 1999, J. Acoust. Soc. Am. 106: 1633-1654; Brown, Stecker & Tollin 2015, J. Assoc. Res. Otolaryngol. 16: 1-28) document the process as continuous and automatic.
NOMN's variation is in the signal before it reaches the speaker, and it changes from moment to moment. A room and a speaker apply the same fixed response to everything. We measured NOMN's layers in the signal itself, and we have not measured how they interact with a given room.
DAC reconstruction
The speaker, room, HRTF and ear canal act on the modulation as part of the audio, and none of them removes it selectively.
Low-depth chorus and filtered noise
The results are measurably different. LFO-driven modulation is periodic, and in our three-minute test it was told from NOMN in 52 of 52 pieces. Random noise is the closer case. We tested the two forms a humanizer uses, independent Gaussian offsets and 1/f noise, each set to NOMN's size, and each was told from NOMN in 52 of 52 pieces. NOMN's activity is organized in time, in bursts of motion followed by calmer stretches. Noise is uniform.
Natural timing variation has structure. Hennig (2014, PNAS 111: 12974-12979) documented long-range correlations in the timing of professional drum performances, consistent with broader work on human motor timing (Gilden, Thornton & Mallon 1995, Science 267: 1837-1839). NOMN's models are trained on the voice, and we have measured that its output is structured in the sense above. We have not yet measured how closely that structure matches any particular natural source, and we do not claim that it does.
MQA and C Wave
MQA is an encoding and decoding format: it changes how a file is stored and needs a licensed decoder for full playback. NOMN does not touch the encoding chain. It adds timing variation to the audio samples before conversion, and the result is an ordinary PCM signal that plays on any equipment. We intend NOMN for integration into hardware and streaming clients and for use by mastering engineers as a final process.
C Wave, as we read its published description, holds that PCM is "non-continuous", that the brain detects the discontinuity, and that a reverb-like process fills the gaps. We do not share that diagnosis. A reverb algorithm running on PCM is still PCM, and the sampling theorem guarantees that properly bandlimited PCM is mathematically equivalent to a continuous waveform up to the Nyquist frequency. Our claim is that digital synthesis and digital playback add no timing variation of their own, where every acoustic source and every analog transport does. That is a claim about sources and transports, and it leaves sampling theory alone.
Humanizer plugins
Humanizers add random variation. NOMN's variation comes from models trained on the voice, and in our measurements it is distinguishable from Gaussian and 1/f random variation of the same size.
Humanizers operate on MIDI event timing before audio is rendered: a humanizer on a quantized MIDI snare moves the hit. NOMN operates on the audio signal and moves it continuously. NOMN also processes finished audio files.
Audibility
Whether a listener would succeed at telling NOMN-processed audio from unprocessed audio in a controlled A/B trial is an empirical question. We intend to investigate it with independent, pre-registered perceptual research, and we will publish the results.
Side effects
Moving audio in time also moves its pitch slightly. At Juiciness 1 the motion is 0.054 % in speed, 0.93 cents RMS, beside 1.18 cents for our tape model. The table under How far NOMN moves the audio gives the figure at each setting. Because of this the processed file will not cancel against the original in a null test, and a steady tone acquires low-level sidebands. Both are measurable, and we report them here.
The name
References
INTERAURAL TIME DIFFERENCE THRESHOLDS
• Klumpp, R.G. & Eady, H.R. (1956). "Some Measurements of Interaural Time Difference Thresholds." Journal of the Acoustical Society of America 28(5): 859-860. 9 µs threshold for band-limited noise, 11 µs for a 1000 Hz tone, 28 µs for clicks (75 % correct discrimination, ten listeners).
• Mills, A.W. (1958). "On the Minimum Audible Angle." Journal of the Acoustical Society of America 30(4): 237-246. Angular acuity in sound localization (about 1 degree near the midline).
• Brughera, A., Dunai, L. & Hartmann, W.M. (2013). "Human interaural time difference thresholds for sine tones: The high-frequency limit." Journal of the Acoustical Society of America 133(5): 2839-2855. Thresholds near 10 µs for pure tones at mid frequencies, with a high-frequency cutoff around 1.4 kHz.
NEURAL CODING OF TEMPORAL STRUCTURE
• Palmer, A.R. & Russell, I.J. (1986). "Phase-locking in the cochlear nerve of the guinea-pig and its relation to the receptor potential of inner hair-cells." Hearing Research 24(1): 1-15. Phase-locking of auditory-nerve fibers and its upper frequency limit.
• Moore, B.C.J. (2008). "The role of temporal fine structure processing in pitch perception, masking, and speech perception for normal-hearing and hearing-impaired people." Journal of the Association for Research in Otolaryngology 9(4): 399-406. Review of temporal fine structure (TFS) and its perceptual role.
• Smith, Z.M., Delgutte, B. & Oxenham, A.J. (2002). "Chimaeric sounds reveal dichotomies in auditory perception." Nature 416: 87-90. Listeners rely on TFS for pitch and localization, and on ENV for speech recognition in quiet.
• Lorenzi, C., Gilbert, G., Carn, H., Garnier, S. & Moore, B.C.J. (2006). "Speech perception problems of the hearing impaired reflect inability to use temporal fine structure." Proceedings of the National Academy of Sciences 103: 18866-18869. TFS and speech perception in noise.
• Hopkins, K. & Moore, B.C.J. (2009). "The contribution of temporal fine structure to the intelligibility of speech in steady and modulated noise." Journal of the Acoustical Society of America 125(1): 442-446.
SOURCE AND ENVIRONMENT SEPARATION
• Wallach, H., Newman, E.B. & Rosenzweig, M.R. (1949). "The Precedence Effect in Sound Localization." American Journal of Psychology 62(3): 315-336. Listeners localize sounds from the first-arriving wavefront.
• Bregman, A.S. (1990). Auditory Scene Analysis: The Perceptual Organization of Sound. MIT Press.
• Litovsky, R.Y., Colburn, H.S., Yost, W.A. & Guzman, S.J. (1999). "The Precedence Effect." Journal of the Acoustical Society of America 106(4): 1633-1654. Review of the precedence effect and echo suppression.
• Brown, A.D., Stecker, G.C. & Tollin, D.J. (2015). "The Precedence Effect in Sound Localization." Journal of the Association for Research in Otolaryngology 16(1): 1-28.
LATENCY AND MUSICAL PERFORMANCE
• Jack, R.H., Mehrabi, A., Stockman, T. & McPherson, A. (2018). "Action-sound Latency and the Perceived Quality of Digital Musical Instruments." Music Perception 36(1): 109-128.
• McPherson, A., Jack, R. & Moro, G. (2016). "Action-Sound Latency: Are Our Tools Fast Enough?" Proc. NIME 2016.
• Schmid, A., Ambros, D., Bogon, J. & Wimmer, R. (2024). "Measuring the Just Noticeable Difference for Audio Latency." Proc. Audio Mostly 2024 (ACM): 325-331. Mean JND of 49 ms at 0 ms base latency and 27 ms at 64 ms.
• Rasch, R.A. (1979). "Synchronization in performed ensemble music." Acustica 43: 121-131.
NATURAL TIMING STATISTICS
• Hennig, H. (2014). "Synchronization in human musical rhythms and mutually interacting complex systems." Proceedings of the National Academy of Sciences 111(36): 12974-12979. Long-range correlations in the timing of professional drum performance.
• Gilden, D.L., Thornton, T. & Mallon, M.W. (1995). "1/f noise in human cognition." Science 267: 1837-1839.
Hearing has the finest temporal resolution of the human senses. Listeners can detect a 9 µs timing difference between the two ears (Klumpp and Eady, 1956). A 60 Hz monitor draws one frame every 16.7 ms, 1,852 times that interval.
Every acoustic source varies in time. A voice, a bowed string and breath across a flute never repeat a cycle exactly. Analog transports vary as well: the mechanical instability of a turntable or tape machine adds slow drift to whatever it plays. Vinyl pressings and analog signal chains carry this drift into the listening room.
Digital synthesis and digital playback add no timing variation of their own. DAWs, digital synthesizers, drum machines, samplers and streaming playback run from a rigid clock, and audiophiles buy 10 MHz external clocks to tighten it further. A synthesized or sequenced source has no timing variation unless a process adds it.
NOMN adds a layer of timing variation learned from the human voice. It works at two scales at once: a slow drift of the whole signal, in the range of tape and record-player motion, and microsecond-scale variation, at the scale of temporal fine structure. The variation is generated live for every moment of audio that passes through, and does not loop.
Tape and record-player emulations model the mechanics of two machines. NOMN's timing models are trained on speech and on electroglottograph recordings of the vocal folds.
NOMN operates in what Marcel Duchamp called the inframince (the infrathin): the separative difference between two things that appear identical, such as two casts from one mold. Here the two things are the timing of a voice and the timing of a clock. Duchamp wrote his notes on the term in the 1930s and 1940s, and they were published in 1980.
How It Works
NOMN is built on generative models of timing trained on the human voice. The drift model learned from multilingual recorded speech. Our research across 103 languages found that timing at this scale differs little from one language to another: in our variance analysis, speaker identity accounted for 25 to 50 % of the variance and language for 3 to 9 %. We read that as support for a common physical basis in the mechanics of the voice. The fine-timing model learned from electroglottograph recordings, in which electrodes on the neck measure the contact of the vocal folds directly, so the training material is the glottal cycle as measured at the larynx. The glottal research points the same way: the fine-timing behavior appeared in all 11 speakers we measured.
At runtime the models produce a continuous stream of timing values, applied to incoming audio in two layers. The drift layer moves the whole signal. The fine-timing layer adds variation at the scale of microseconds. Juiciness sets the depth of both layers. The fine-timing amount scales the fine-timing layer.
The pattern is the model's own and is the same whatever audio is playing; it is generated continuously and does not repeat within any session we have measured.
Compared With Existing Modulation
Earlier timing modulation takes three forms. An LFO drives a chorus or a vibrato. Tape and record-player emulation reproduce the motion of a transport. A humanizer uses random numbers. NOMN can be told from each of them by measurement. We put one steady synthesized tone, three minutes long, through each process and read its timing from the tone's phase, which is the method used to measure wow and flutter. One of the processes is a tape emulation you can download: the open-source CHOW Tape Model plugin, run with its wow and flutter controls at half depth.
Three minutes of each. Timing offset of the same tone, every process drawn at the same size on the same scale.
Speed-deviation spectrum. The same eight processes as a spectrum of speed deviation, 0.2 to 30 Hz, on one scale.
Listen
Six short loops from our own demo tracks. Each can be switched between the untreated source, NOMN at stated settings, and the tape model and record player model, and switching keeps your place in the music. Every version is matched in level and aligned in time.
Use Cases
Production
Synthesizers, drum machines, samplers and sequenced parts: sources that have no timing variation of their own.
Mastering & Post-Production
A time-domain process for the mastering chain, used alongside EQ, compression, spatial processing and loudness control.
Licensing
Available for integration into hardware, software and streaming infrastructure. Ask us: hello@nomn.jp.
Scope
NOMN is not an equalizer, a compressor or a spatial processor. It does not change level or loudness and adds no noise or saturation. The modification is in the time domain, and a change in time is also a small change in pitch: 0.93 cents RMS at Juiciness 1. A processed file will not null against its original.
We do not claim that digital audio is missing temporal fine structure, or that NOMN restores something the format lost. A digital recording carries the fine structure of what was recorded. NOMN adds a layer of its own.
Formats & Access
Desktop app: in development. Late 2026.
Patent Status: Patent pending (Japan, 2026). POLYTOPE KK.
Digital audio basics
Digital audio is harder to picture than pixels on a screen, and accounts of it in forums and audiophile communities disagree widely. Reproduced sound is a perceptual illusion built from numbers, and it works far better than its mechanism suggests.
Recording, or rendering an audio file, takes fast measurements of a continuously varying signal and stores them as a sequence of numbers.
A sample is one measurement: an integer or floating-point number giving the amplitude of the waveform at one instant. A file is a dense series of amplitude values over time.
The sample rate is the number of measurements per second. CD audio stores 44,100 per second (44.1 kHz); 96 kHz stores 96,000. A correctly bandlimited sampled signal represents the original waveform completely up to the Nyquist frequency, so nothing is missing between samples. The DAC's reconstruction filter produces the one continuous waveform the samples define. Rates above 48 kHz move the filter's transition band further above the audible range.
DSD (Direct Stream Digital) runs from 2.8224 MHz (DSD64) to 22.5792 MHz (DSD512). It stores 1 bit per sample, and noise shaping at the high clock rate supplies the resolution the single bit lacks.
A sample carries no information about frequency, timbre or pitch. Those properties exist in the pattern across many samples. A speaker cone needs a position for each instant, and a sequence of positions traces any waveform within the band. The cone displaces air, and the displacement travels as a compression wave.
To represent a frequency F, a system must sample at more than 2F. Hearing extends to 20 kHz in young listeners, so 44.1 kHz and 48 kHz cover the audible band with margin. At 343 m/s (air at 20 °C), a 20 kHz wave is 17.2 mm long.
44.1 kHz sampling → 22.05 kHz max frequency → 15.6 mm wavelength
48 kHz sampling → 24 kHz max → 14.3 mm
96 kHz sampling → 48 kHz max → 7.1 mm
192 kHz sampling → 96 kHz max → 3.6 mm
384 kHz sampling → 192 kHz max → 1.8 mm
768 kHz sampling → 384 kHz max → 0.89 mm
DSD is not directly comparable on this table. Its clock rate at DSD512 is 22.5792 MHz, which is the rate of a 1-bit modulator, and its usable audio bandwidth is set by the noise-shaping filter.
Bit depth sets the number of possible amplitude values per sample. 24-bit gives 16,777,216 values, which sets the dynamic range and the noise floor. Bit depth sets resolution in amplitude only.
A buffer is a block of consecutive samples that software, drivers and hardware pass as a unit, typically 64, 128 or 512 samples. At 96 kHz a 64-sample buffer holds 0.667 ms of audio. Smaller buffers lower latency (the time from input to output) and raise CPU load and the risk of dropouts caused by the operating system, drivers or firmware. Larger buffers reduce CPU load and add delay, which matters in live performance and monitoring.
With more than one channel, each instant has one sample per channel, and that group is a frame. Stereo at 96 kHz produces 96,000 two-sample frames per second. Buffer sizes are counted in frames, so a buffer size corresponds to a fixed duration at any channel count.
The audio engine fills each buffer, processes it and passes it to the DAC. At 96 kHz with 64-frame buffers this hand-off happens 1,500 times per second. The DAC converts the numbers to a continuous voltage that drives the speaker cone to the positions the samples describe. The output is continuous at any buffer size, because reconstruction produces a continuous waveform.
Subtlety
At low Juiciness the effect is subtle. It changes how audio moves in time and does not register as a discrete change in the way an EQ boost does. Whether it matters for a given listener, recording or playback chain is an empirical question, and we have not yet run the listening tests that would answer it.