Bias Tape: a multitrack tape machine model
Abstract
A musician who wants tape on every track usually loads a tape plugin on every channel and gets a separate machine on each one, each with its own speed error. Bias Tape is a real-time model of an analog tape machine that treats the whole session as one machine. It runs in Bias, our Mac app for TASCAM Model mixers, and on the Model 2400 up to 22 tracks pass through Bias Tape at once. Each track runs a physically motivated chain of standards-based record emphasis, a Jiles–Atherton magnetization model (the standard physical model of how magnetic material saturates) solved at twice the sample rate, complementary playback de-emphasis, playback-head losses, head bump, wow and flutter, and hiss. Our main contribution is treating the tracks of a session as one machine. Speed deviation is a pure function of the shared transport clock, so every track drifts sample-identically. Tape speed, formulation, format and noise belong to the machine and follow across tracks. A shared headstack lets each track pick up its neighbors at a level set by the format's track width.
We tested the model against predictions we registered before measuring. Third-harmonic distortion reaches 3% within 0.2 dB of the target level for three tape formulations. One mono channel costs 0.62% of real time on an M-series laptop. We report as failures the three predictions that did not hold as registered. The CPU cost missed its bound by 3%, and two predictions about level-dependent treble loss also failed. We have not compared the model with a physical machine.
1. Introduction
Developers usually build and judge tape emulation one channel at a time, so a plugin models one track of one machine. A musician who wants tape everywhere loads an instance on every channel. Each instance runs as an independent machine with its own uncorrelated speed error and its own settings, and no signal passes from one copy to another. On a real multitrack recorder, one capstan (the spindle that pulls the tape) moves one tape past one headstack (the block of record and playback heads) for every track.
Bias Tape runs in Bias, our Mac app for TASCAM Model mixers (Model 12, 16, 24 and 2400). The Model mixers are analog desks with a built-in multitrack recorder and a multichannel USB interface. Bias takes each channel's send over USB and returns it to the desk through a chain of Audio Units, the plugin format macOS uses. Each strip has a dedicated tape slot. On the Model 2400, whatever occupies the tape slot runs on up to 17 strips (12 mono, 5 stereo) inside one audio callback, with a round trip of about 13 ms at a 128-sample buffer. That's why the design bounds the cost per channel and rules out allocation or locks on the audio thread. Bias Tape must also report its latency and must never leave a filter ringing on a channel with no source connected.

Licensing sets a further constraint. The best-known open tape model, CHOW Tape Model, is distributed under the GPLv3, a license that does not suit a closed commercial host like Bias. So we wrote Bias Tape from the published physics and from measurement papers, without reference to the source of any existing implementation.
This paper makes three contributions:
- A per-track signal path in which treble headroom follows from the playback equalization standards, using an exactly complementary emphasis pair around the magnetization model.
- A multitrack formulation that makes speed deviation a pure function of the transport's sample clock, shares machine-level settings across linked tracks and carries adjacent-track bleed through a headstack at a level set by track width. The result is independent of render order.
- An evaluation in which we registered every claim as a prediction before measuring it, and report the predictions that failed beside the ones that held.
2. Related work
Bias Tape draws on four groups of prior work on tape.
Hysteresis models. Jiles and Atherton's 1986 model of ferromagnetic hysteresis describes magnetization as an anhysteretic curve that the material follows imperfectly, with domain-wall pinning producing the loop. The Jiles–Atherton model is the standard physical account and the basis of most analog tape models in audio. Chowdhury's 2019 DAFx paper applies the hysteresis model to tape with fourth-order Runge–Kutta at 16 times oversampling and simulates the high-frequency bias signal explicitly. Bias Tape solves the same equation by a different method. The code follows the 1986 formulation and was written without consulting any other implementation.
Head and medium losses. Wallace's 1951 analysis of magnetically recorded signals gives the losses that dominate the treble: separation loss falling exponentially with the ratio of head-to-tape spacing to wavelength, thickness loss from the depth of the recorded layer, and gap loss from averaging across the reproduce gap. All three losses are wavelength effects. A given head is therefore brighter at higher tape speeds. Bias Tape uses the textbook expressions directly.
Measured and data-driven models. Mikkonen and colleagues' 2023 DAFx work measures a consumer reel-to-reel machine and fits a neural model to the measurements. The same paper reports the real transport's delay-modulation depth, the frequency band that the modulation occupies and a noise floor that rises toward the low end. The wow, flutter and hiss in Bias Tape draw on those reported figures. Arnardottir, Abel and Smith's 2008 AES model of the Echoplex tape delay is an earlier reference for treating capstan and pinch-wheel irregularity as a quasi-periodic modulation of delay. Bias Tape borrows the reported magnitudes and shapes without fitting to any measurement set.
Equalization standards and bias. Adriaensen's 2025 Linux Audio Conference paper on tape simulation argues that the treble behavior of a tape machine follows from the recording and playback equalization standards and a nonlinearity. The paper also lists the playback time constants in use as NAB's 50 µs, IEC's 70 µs at 19 cm/s and 35 µs at 38 cm/s, AES's 17.5 µs at 76 cm/s and the 120/70 µs pair for cassette. McKnight's work on biasing describes what AC bias does to linearity. Bias Tape relies on the equalization standards more than on any other source, taking its treble headroom from the pre-emphasis that the standard sets for each speed and tape.
Every work cited here models a single channel. We have not found the multitrack formulation of Section 7 anywhere in the audio literature. The underlying physics is standard recording practice.
3. The signal path
One channel runs eleven stages, in the order a signal meets them on a real machine.
- Inputsanitize and drive
- Recordemphasis
- 2×upsample
- Jiles–Athertonmagnetization
- 2×downsample
- Playbackde-emphasis
- Head lossesand bias tilt
- Head bump
- Wow andflutter
- Neighborbleed
- Hiss
- Output
The order matters in three places. Emphasis comes before the nonlinearity and de-emphasis after it. Treble therefore meets the tape hotter than the bass does and saturates sooner. That's why slow tape runs out of treble headroom first. Oversampling brackets only the nonlinearity, the one stage that creates new frequencies. Wow and flutter come after the filters and before the noise, because a real machine's speed error modulates the recorded signal and leaves the electronics' own hiss untouched.
Drive and output are ramped across each block to keep a moved control from stepping. Input is clamped to ±12 dBFS. Any value that is not a finite number is replaced by zero before reaching the tape, since the feedback in the hysteresis model and the recursive filters would otherwise carry a bad sample forever.
An idle strip uses very little CPU. After one second below −90 dBFS, the noise floor of a desk with an open input, every filter tail in the chain has decayed. The channel then parks. A parked channel clears its delay line and outputs only hiss and its neighbors' bleed until signal returns.
4. Magnetization
The tape itself is one Jiles–Atherton state per channel, with a dimensionless magnetization M driven by a dimensionless field H. Saturation magnetization and the Langevin shape parameter are both set to 1. M therefore lies in −1 to 1, with H measured in units of the anhysteretic knee. The effective field is He = H + αM with α = 0.02. The Langevin function gives the anhysteretic curve.
δ is the sign of dH/dt. The irreversible term is dropped whenever δ(Man − M) is negative, the usual correction that keeps susceptibility from going negative at a turning point. The denominator of the irreversible term is floored at a quarter of (1−c)k so the slope stays finite. The Langevin function and its derivative are tabulated over 0 to 16 in 4096 points with linear interpolation, backed by a series expansion near zero and the 1 − 1/x asymptote beyond the table. Any input that is not a number falls into the asymptote branch and cannot reach the table index.
Two parameters carry the user-facing controls.
Formulation sets k, the pinning constant, which is the loop's width. The three tape types take values of 0.5, 0.65 and 0.9, running from brighter and more forgiving to harder and cleaner.
Bias sets c, the reversible fraction. Mapping bias to c is the paper's one interpretive move. AC bias makes tape linear. In Jiles–Atherton terms, a well-biased tape follows its anhysteretic curve almost exactly and saturates only at the top. Aligned bias is therefore c = 0.96. Under-biasing opens the loop steeply and makes quiet signals gritty, with c falling to 0.40 at the extreme. Over-biasing has little room left and takes c only to 0.99, leaving the treble tilt described in Section 5 to carry the audible change. AC bias is not modeled as a high-frequency carrier added to the signal. A carrier would demand far higher oversampling for no audible gain at the sample rates in use.
The equation is solved with midpoint Runge–Kutta at twice the strip's sample rate. A two-times half-band stage with a 47-tap Kaiser-windowed sinc brackets the solver. Chowdhury oversamples by a far larger factor of 16 because his model simulates the bias carrier itself. Representing bias as reversibility removes the need for that much oversampling. Bias Tape runs at the lower rate to save CPU.
Absolute field has no meaning until a reference fixes it. Each configuration is therefore calibrated once, off the audio thread. A bisection of 18 geometric steps finds the field at which third-harmonic distortion of a 1 kHz tone reaches 3%. Full scale is then placed at the formulation's rated level for that figure, which for the three types is −5, −4 and −1 dBFS. Distortion is judged after playback, at the point where a listener hears the signal. The probe tone is therefore lifted by the record emphasis at 1 kHz, with its third harmonic weighted by the emphasis ratio between 1 kHz and 3 kHz. A second tone 30 dB down then sets a makeup gain that returns quiet signals at unity gain. Calibration is skipped for settings that cannot move it, such as wow depth or output level.
5. Equalization and the playback head
The playback equalization standard for the speed and tape in use sets the treble headroom, because the standard determines how much hotter the treble goes onto the tape.
| Speed | Standard | Type I | Type II / IV |
|---|---|---|---|
| 1⅞ IPS (4.76 cm/s) | IEC cassette | 120 µs | 70 µs |
| 3¾ IPS (9.5 cm/s) | double-speed cassette | 70 µs | 50 µs |
| 7½ IPS (19 cm/s) | NAB | 50 µs | 50 µs |
| 15 IPS (38 cm/s) | IEC | 35 µs | 35 µs |
| 30 IPS (76 cm/s) | AES | 17.5 µs | 17.5 µs |
The record side applies a first-order shelf with the time constant from the table, limited to a gain of 6.3 (16 dB) so that the boost plateaus:
The shelf is bilinear-transformed to a one-pole, one-zero section. Playback applies the same section with numerator and denominator exchanged. The pair is exactly complementary by construction. A quiet signal therefore comes back unaltered whatever the time constant. In testing, a pair of truncated FIR filters failed to cancel and left 15 IPS 2.3 dB down at 18 kHz. The IIR section serves both the record and playback sides instead.
The playback-head loss model uses Wallace's separation and gap terms, with head geometry switched by speed. Cassette speeds use 0.2 µm spacing and a 1.0 µm gap, against 0.5 µm spacing and a 2.5 or 4.0 µm gap at the studio speeds. Coating thickness loss is set to zero, because the equalization standard exists to make up for that loss. Including thickness loss in the model as well would take the treble out twice.
Since the wavenumber k depends on tape speed v, one head is brighter at 30 IPS than at 1⅞ IPS without any parameter changing. Two more terms ride the same filter. A tilt of −4 dB per unit of bias above 10 kHz makes under-biased tape brighter and lets over-biased tape erase its own treble. A second term corrects the droop of the half-band filters near the top of the band. The head losses, the bias tilt and the droop correction are multiplied and realized as one 64-tap minimum-phase FIR. The FIR is designed by folding the real cepstrum of the target magnitude through a 1024-point DFT, with the last quarter of the taps faded by a raised cosine.
Head bump is two biquads. A +2 dB peak sits at 4.5 times the speed in IPS, which is 17 Hz at 3¾ IPS and 135 Hz at 30 IPS. A second-order high-pass at 45% of the peak frequency also removes the DC offset left by a magnetized tape.
6. Transport and noise
The model treats speed error as a displacement of the tape. The displacement from the nominal tape position is computed in seconds, as a closed-form function of absolute transport time:
W and F are the peak speed deviations the Wow and Flutter controls set. Wow uses components at 0.55 Hz and 1.37 Hz plus a slow random walk. Flutter combines the capstan rotation rate, a partial at 2.63 times that rate and a faster random term. Capstan rate is taken as 5.8 to 24.2 Hz across the five speeds. Each division by 2πf is the integration from speed error to displacement. That's why wow moves pitch far more than flutter at the same percentage. The random terms are a hash of the sample index with smoothstep interpolation, which makes them, too, a function of t alone with no running state.
Because every term depends only on t, the transport can be shared, with every instance on the machine evaluating the same e(t) from the same host timestamp. All tracks therefore drift together, sample for sample, without communicating. The scheme needs no leader track and no shared buffer of speed values. In Section 9, two stages on one clock produce bit-identical output.
The excursion is read from a delay line by four-point Lagrange interpolation, with a base delay of the peak excursion plus three samples so the read point can swing both ways. The base delay is added to the reported latency and glides to its new value when the depth control moves.
Hiss is white noise from an xorshift generator plus the same noise through a one-pole low-pass at 150 Hz at half the amplitude, with the sum normalized to unit RMS. The resulting spectrum is flat with a rise at the low end, as in the noise of a measured machine. Hiss level falls 3 dB for each doubling of tape speed and 3 dB for each doubling of track width. The Hiss control therefore reads as a figure for one cassette track at 3¾ IPS. For the same setting, the model puts a 2-inch 24-track machine at 30 IPS about 12 dB quieter.
7. One machine, many tracks
Every Bias Tape instance in Bias is one track of a shared machine. The channel strips record onto a multitrack, which is mixed down to a second machine, the mixdown deck. When the host loads a track, the track joins a machine. A Link switch on each track takes the track off the machine and gives it its own settings.
Each setting sits with the physical part it controls, either the machine or the track. Format, Speed, Tape, Wow, Flutter and Hiss belong to the machine. A change to a machine setting on any linked track reaches every linked track, because one transport cannot run at two speeds and one reel cannot hold two formulations. Input, Bias and Output belong to the track, because alignment and gain staging are set per channel on a real machine too. A track that arrives with saved state sets the machine from that state. A fresh track adopts the machine's current settings. The machine holds its settings even with no track loaded. The next track to load finds the machine as it was left.

Format chooses the tape and the width of each track on it. Track width in turn fixes two things the user never adjusts directly.
| Format | Track width | Neighbor bleed at 1 kHz |
|---|---|---|
| Cassette 4-track | 0.6 mm | −45 dB |
| ¼″ 8-track | 0.5 mm | −45 dB |
| ½″ 8-track | 1.0 mm | −50 dB |
| 1″ 16-track | 1.0 mm | −50 dB |
| 2″ 16-track | 1.8 mm | −60 dB |
| 2″ 24-track | 1.1 mm | −55 dB |
| ¼″ 2-track | 2.0 mm | −55 dB |
| ½″ 2-track | 5.0 mm | −60 dB |
Narrow tracks are noisier and leak more. A wide tape carrying 24 tracks gives each track less width than the same tape carrying 16. That's why the 2-inch 24-track is noisier and leakier than the 2-inch 16-track.
The headstack
Tracks hear their neighbors through a shared structure indexed by the transport's sample count. The structure holds 24 lines of one block each, plus the write position of each line. A track writes its finished block at position t and reads the blocks of tracks above and below it covering [t − n, t), one block old.
The one-block lag makes the result independent of render order. The host does not guarantee which strip's callback runs first. Within the same block, a read could return a neighbor's current block or that neighbor's previous block depending on callback order, which would make the output non-deterministic. Reading strictly one block behind gives every track the same input however the callbacks are scheduled. The lag adds one buffer of delay on the bleed path, which at 128 samples is under 3 ms and inaudible against a signal already 45 dB down.
The bleed is the sum of the neighbors, passed through a gain plus three times a one-pole low-passed copy at 200 Hz and normalized so that the table's figure is the level at 1 kHz. Crosstalk on a real headstack rises toward the low end, where wavelengths are long enough to reach across the guard band. In Bias Tape the bleed is about 10 dB stronger at 60 Hz than at 1 kHz. A silent track still plays its neighbors' bleed, as on a real machine. An unlinked track neither sends nor receives.
On a stereo strip the two channels occupy adjacent tracks. Left and right therefore bleed into each other at the format's level.
8. Implementation
The model is about 900 lines of Swift with Accelerate, split across two files. The DSP file depends on nothing but Accelerate and can be measured offline. An Audio Unit wrapper fills the second file.
The wrapper is an AUAudioUnit subclass registered into the engine's own process at startup, with a component type of its own. The subclass is not installed on the Mac and stays invisible to every other application. The host already drives plugins through the version-2 render interface and handles parameter trees, factory presets, full-state save and restore, metering and editors through that one path. Registering in-process puts the built-in model on exactly that path, which keeps the host uniform and free of special cases. The model appears in the plugin list beside the Audio Units installed on the Mac, loads into a slot the same way and saves its state with the song.
Real-time safety is handled by keeping everything the render thread touches fixed. Buffers are allocated when the unit is prepared, sized for the largest block. Anything derived from a setting, including the calibration bisection, the minimum-phase FIR design and the biquad coefficients, is computed off the audio thread into an immutable configuration object. The render callback picks up a new configuration by swapping an unmanaged pointer at the top of a block. Superseded configurations are held in a list for several seconds before release. The audio thread therefore cannot dereference freed memory, even if a control moves while the thread is mid-block. Parameters that cannot affect calibration reuse the previous calibration outright.
Parameters fall into two groups, Track and Machine, matching the division in Section 7. The grouping lets the host's generic editor draw the distinction without knowing anything about tape. The unit has ten parameters, four in Track (Input, Bias, Output and Link) and six in Machine (Format, Speed, Tape, Wow, Flutter and Hiss). The unit also carries six factory presets and a state dictionary that holds the settings and the current preset.
Reported latency is 23 samples from the two half-band filters, plus the wow-and-flutter base delay when either is non-zero. With the default 0.06% wow and 0.04% flutter at 44.1 kHz, the total latency is about 0.8 ms.
9. Evaluation
We wrote every claim below as a prediction, stating the figure it had to hit and what would count as failure, and lodged each prediction in a signed record before running the measuring code. We report results against the registered predictions, including the three that did not hold.
Two harnesses measure the model. The first harness drives the DSP class directly, offline and deterministically. The second loads the model as the host loads any plugin, through the version-2 render interface, and checks that path. Figures below come from a single run of both harnesses on an M-series laptop at 44.1 kHz with a 128-sample block. We have not measured on Intel Macs.
Single-value claims
| Measurement | Predicted | Measured | Held |
|---|---|---|---|
| Gain at 1 kHz, −30 dBFS | within 1 dB | −0.12 dB | yes |
| 3% third harmonic, Type I | −5 dBFS ± 1.5 | −4.88 dBFS | yes |
| 3% third harmonic, Type II | −4 dBFS ± 1.5 | −3.84 dBFS | yes |
| 3% third harmonic, Type IV | −1 dBFS ± 1.5 | −0.80 dBFS | yes |
| Wow depth at 0.2% setting | 0.15–0.25% | 0.188% | yes |
| Two stages on one clock | bit-identical | identical | yes |
| Hiss at −60 setting | −60 dBFS ± 1.5 | −59.95 dBFS | yes |
| Output under abuse | finite, under +12 dBFS | finite, +5.29 dBFS peak | yes |
| Settled DC after a burst | under −80 dBFS | −107.5 dBFS | yes |
| Latency, transport steady | 23 samples ± 2 | 23 samples | yes |
| Cost, one mono channel, 10 s | under 60 ms | 61.8 ms | no |
| Host render path vs direct | identical samples | 0 of 44,032 differ | yes |
| State round trip | values and preset equal | equal, 591 bytes | yes |
The cost claim failed by 3% on a computer that was compiling two applications during the run. Earlier runs with no other load measured 57 to 58 ms. Because the registered run came in at 61.8 ms, we record the claim as failed. At 0.62% of real time per mono channel, 22 channels of tape on the Model 2400 would take about 14% of one core.
Frequency response
- 1⅞ IPS
- 3¾ IPS
- 15 IPS
- 30 IPS
Data
| Hz | 1⅞ IPS | 3¾ IPS | 15 IPS | 30 IPS |
|---|---|---|---|---|
| 30 | +0.2 dB | +0.6 dB | −2.7 dB | −12.4 dB |
| 50 | +0.2 dB | +0.3 dB | +0.8 dB | −4.8 dB |
| 70 | +0.2 dB | +0.2 dB | +1.9 dB | −1.4 dB |
| 100 | +0.2 dB | +0.1 dB | +1.0 dB | +0.7 dB |
| 200 | +0.2 dB | +0.1 dB | +0.2 dB | +1.0 dB |
| 500 | +0.1 dB | 0.0 dB | 0.0 dB | +0.1 dB |
| 1k | 0.0 dB | 0.0 dB | 0.0 dB | 0.0 dB |
| 2k | −0.2 dB | −0.1 dB | −0.1 dB | 0.0 dB |
| 4k | −0.7 dB | −0.3 dB | −0.2 dB | −0.1 dB |
| 6k | −1.1 dB | −0.5 dB | −0.4 dB | −0.2 dB |
| 8k | −1.6 dB | −0.7 dB | −0.5 dB | −0.4 dB |
| 10k | −2.3 dB | −0.9 dB | −0.7 dB | −0.6 dB |
| 12k | −3.0 dB | −1.2 dB | −1.0 dB | −0.7 dB |
| 14k | −3.8 dB | −1.5 dB | −1.2 dB | −1.0 dB |
| 16k | −4.7 dB | −1.9 dB | −1.5 dB | −1.2 dB |
| 18k | −5.8 dB | −2.3 dB | −1.8 dB | −1.3 dB |
| 20k | −7.2 dB | −3.1 dB | −2.4 dB | −1.5 dB |
tapecheck T3 · 4 speeds × 17 frequencies
The four curves differ only in tape speed, with no filter parameter changing from one curve to the next. Slow tape loses treble because head losses are wavelength effects. The equalization standard for each speed also sets how much treble went onto the tape. At 1⅞ IPS the response is 3 dB down at 14 kHz. The 15 and 30 IPS curves hold to 20 kHz. The same speed dependence appears at the low end, where head bump sits at 4.5 times the speed. The bump is a gentle lift around 17 Hz at 3¾ IPS and a pronounced 1 dB shelf near 200 Hz at 30 IPS. The machine's own high-pass then removes the lowest frequencies.
Level dependence, and what broke
- 1 kHz
- 10 kHz
Data
| dBFS | 1 kHz | 10 kHz |
|---|---|---|
| −30 | 0.0 dB | +0.8 dB |
| −24 | +0.1 dB | +1.2 dB |
| −18 | +0.1 dB | +1.2 dB |
| −12 | +0.1 dB | −0.6 dB |
| −6 | −0.6 dB | −5.5 dB |
| 0 | −2.6 dB | −13.1 dB |
| +6 | −6.4 dB | −21.5 dB |
tapecheck level sweep · 7 levels × 2 frequencies
The registered prediction said that at 3¾ IPS a 10 kHz tone at −10 dBFS would compress at least 3 dB more than a 1 kHz tone at the same level. At 30 IPS the two tones were to stay within 1.5 dB of each other. Neither part of the prediction held. Measured at −10 dBFS, the difference at 3¾ IPS is 1.97 dB, short of the 3 dB claimed. At 30 IPS the two tones differ by 1.67 dB, just outside the 1.5 dB bound and in the wrong direction.
The level sweep in the chart explains the failure. Above −6 dBFS the model behaves as predicted. At 0 dBFS the 10 kHz tone is 13.1 dB down against 2.6 dB for the 1 kHz tone, a 10 dB spread. At lower levels, between −30 and −18 dBFS, the hysteresis model expands treble by about 1 dB. The crossover from expansion to compression therefore happens later than the prediction assumed. A measurement taken at −10 dBFS catches the model mid-turn. The artifact, which comes from the Jiles–Atherton parameters at the reversibility in use, is the clearest open problem in the model.
The machine
A second track loading onto a machine took all six machine settings from the track already there and kept its own Input setting. Setting speed on either track changed both. With a tone on one track and silence on the other, the silent track's output measured 42.8 dB below the loud track at 1 kHz on Cassette 4-track, against a table figure of 45 dB. The loud track's own −2 dB Output setting accounts for the difference. At 60 Hz the silent track measured 33.3 dB down, showing the low-end rise. Switching the machine to 2-inch 24-track from either track moved both tracks. Bleed then fell to −52.8 dB. Unlinking a track dropped its bleed to −330 dB, which is silence. The unlinked track held its own speed while the other track's speed changed.
10. Limitations
No reference machine. We have not compared the model with a physical tape recorder. The calibration targets for distortion level by formulation, equalization time constants, head spacings, gap widths and crosstalk by format all come from published standards and textbook values. We did not measure a deck. Every figure in Section 9 shows that the model meets its own targets.
The mid-level treble artifact. Between −30 and −18 dBFS the model expands treble by about 1 dB, which broke both halves of the level-dependence prediction. The likely cause is the interaction of the reversibility parameter with the floor on the irreversible term's denominator, which was added for numerical stability and has an audible cost. Fixing the expansion without losing the distortion calibration is the main open problem.
Bias is an abstraction. The model represents AC bias as a reversibility parameter and a treble tilt, without a high-frequency carrier. The model therefore misses the direct effect of bias on short wavelengths. The combined effect of the two controls only approximates the trade-off between bias level, sensitivity and maximum output.
Absent effects. Print-through, dropouts, azimuth error, scrape flutter and the difference between the record and playback heads are all unmodeled. The head difference matters most for the machine, because without a sync and repro distinction the model cannot represent monitoring off the record head during a bounce.
Crosstalk is simplified. Bleed reaches only immediate neighbors, is symmetric and has one frequency shape. A real headstack couples further than one track and behaves differently in record than in playback.
The mixdown machine is unreachable. The model provides for a second machine for the stereo mix. The hardware's main mix has no USB return, which means no track can load onto the mixdown deck today.
11. Future work
Our near-term work has three parts:
- Resolve the mid-level treble artifact, then register the level-dependence predictions again.
- Add a sync and repro head switch that lets the model represent bounce and generation loss. A bounced track already passes through the model twice.
- Add dropouts and creases common to all tracks, which the shared headstack already supports. A crease hits every track at the same instant.
The measurement work needs access to a real machine. Frequency-response and distortion captures from one deck would turn most of Section 9's self-consistency checks into comparisons with hardware. A flutter capture would check the transport against a real machine.
12. Conclusion
The per-channel model combines Jiles–Atherton magnetization and Wallace's head losses with a delay-modulated transport, all established methods. Two choices in the arrangement may be useful in other tape models. Taking treble headroom from the playback equalization standard, through an exactly complementary emphasis pair around the nonlinearity, gives level- and speed-dependent treble behavior without a filter tuned by ear. Calibrating absolute field by bisecting to a rated distortion figure ties each tape formulation to a published target.
We have not found the multitrack formulation elsewhere. Making speed deviation a closed-form function of the shared transport clock gives correlated drift across every track at no extra cost and with no communication between instances. Indexing the shared headstack by transport sample count and reading one block behind makes crosstalk deterministic under any render order. Tying track width to the tape format sets both noise and crosstalk from the Format control.
We have not compared Bias Tape with a physical tape machine. The measurements in Section 9 show that the model matches its own targets.
Disclosure
We produced the model, its measurement harnesses and this paper with Claude, an AI system made by Anthropic, working under Brian Gilmore's direction in a single extended session. Brian Gilmore set the goals and made the design decisions, including treating many tracks as one machine. Claude wrote the code and text, which Brian Gilmore reviewed. We kept the prediction-before-measurement records in a signed Symbia Labs catalog. We did not consult the source of any existing tape plugin.
References
- D. C. Jiles and D. L. Atherton, "Theory of ferromagnetic hysteresis," Journal of Magnetism and Magnetic Materials, vol. 61, pp. 48–60, 1986. doi:10.1016/0304-8853(86)90066-1
- R. L. Wallace, Jr., "The reproduction of magnetically recorded signals," Bell System Technical Journal, vol. 30, no. 4, pp. 1145–1173, October 1951. index
- J. Chowdhury, "Real-time physical modeling for analog tape machines," Proc. 22nd International Conference on Digital Audio Effects (DAFx-19), 2019. pdf
- O. Mikkonen, A. Wright, E. Moliner and V. Välimäki, "Neural modeling of magnetic tape recorders," Proc. 26th International Conference on Digital Audio Effects (DAFx 2023), 2023. arXiv:2305.16862
- S. Arnardottir, J. S. Abel and J. O. Smith III, "A digital model of the Echoplex tape delay," Audio Engineering Society 125th Convention, paper 7649, 2008. AES
- F. Adriaensen, "Tape simulation," Linux Audio Conference (LAC 2025), INSA Lyon, June 2025. Read from a copy of the paper; no public link found.
- J. G. McKnight, "Biasing in magnetic tape recording," 1967. Publication details not verified.