OpenVinyl: 3D-printed records

OpenVinyl turns digital audio into groove geometry on records printed on a hobby FDM printer, played back on a turntable I designed with two collaborators. v1 worked, barely. This month I went back to first principles and reversed its central design decision. v2 is still a work-in-progress, and I plan to print it and run tests in October 2026.

The pivot

v1 encoded audio as groove depth because an FDM bead is 0.48 mm wide, so a lateral wiggle looked impossible at 10 to 20 times the width of a real vinyl groove. It assumed the geometry would go through a slicer, and a slicer can only reproduce what a mesh describes at bead resolution. If the encoder writes the G-code itself, the bead can be placed to about 0.01 mm even though it is 0.48 mm wide.

So v2 is a lateral cut, emitted directly as G-code, on the printer’s fast and precise XY axes. Layer-height quantization, slicer segment merging, and the staircase noise all disappear; the ceramic cartridge works in its designed direction, and the groove pass prints in about 25 minutes instead of hours. The constraint was in the FDM slicer. v2 goes one level lower, so printer precision, and FDM physics are constraints.

What v1 got right

The contact mechanics, the slope limit and the habit of deriving everything from two hardware numbers all carry over unchanged.

• Hertzian contact. An 18 µm LP stylus at 4 g on PLA peaks near 700 MPa, about 14x the yield stress. First play plastically sets the groove. A 3 mil (76 µm) 78-RPM tip cuts peak pressure by about 2.6x, to about 270 MPa, because pressure scales as . That is still over 5x yield. Fortunately this is less of a concern now that this deformation doesn’t act on the geometry encoding the audio.
• Slope constraint. . At 78 RPM, mm and 45°, this is mm·Hz.
• Radius-dependent bandwidth. Outer turns get 3x the inner turns on a 10 inch disc.
• The compressor to fit the amplitude budget, and the three-tier validation: melody recognition, SNR, pitch accuracy.

Slicer constraint based assumptions in v1

Half of v1’s parameters were consequences of the slicer. Each one was re-derived or dropped.

v1 item	v2 verdict	Reason
Z modulation	Dropped	See above. Also, Z on the P1S is the heavy bed on leadscrews; modulating it mid-print would force ~3 mm/s feedrates.
4-bit depth from 0.08 mm layers	Gone	Nothing is quantized. Dynamic range is set by XY repeatability, assumed 0.01 mm, so about 20 dB at mm.
Slicer minimum segment length	Closed by construction	Segment length is a parameter of the emitter, 0.1 mm default.
300 Hz highpass	To be re-derived by listening	It came from “no harmonics inside 681 to 2042 Hz”. In a lateral cut, low frequencies are the easy ones, since slope scales with . Any highpass now exists for amplitude budget: bass eats amplitude, and pitch pays for . Default drops to 100 to 150 Hz.
	Kept as a parameter, number unearned	In lateral the limit is probably the stylus riding up the bead flank plus cantilever resonance, not a tangent angle. Measured by the amplitude-ladder disc.
Nyquist from steps per revolution	Replaced by nozzle smear	A feature shorter than one bead width along the path is averaged out: . Same numbers v1 got, 680 Hz inner and 2 kHz outer, because it is the same physics. A 0.2 mm nozzle doubles it.
Staircase noise at	Gone	No staircase in a planar groove. New periodic noise source: bead-edge ripple from extrusion pulsing, unmeasured.
Cartridge as a given	Now a design variable	Ceramic stays because it is displacement-proportional, so no RIAA and no pre-emphasis. Tip radius and compliance can be chosen for a 0.5 mm crease.
Constant RPM	CAV first, CLV designed for	See below.
Fixed pitch	Variable pitch later	Cutting lathes vary spacing with the signal envelope. We write the G-code, so we can too.
“12 inch” 254 mm disc	It is a 10 inch	A 12 inch LP is 302 mm and does not fit a 256 mm bed.

The v2 groove

Two beads per turn, 1.1 bead widths apart, so a V-crease forms between them. Both beads wiggle together and the stylus rides the crease.

Both beads follow the same radial offset on top of a flat sliced disc. The stylus self-centres in the crease and reads as lateral displacement, which is what a ceramic cartridge is built to sense.

Pitch = bead spacing + bead width + + land. Numbers from the encoder’s own info command:

Quantity	10 inch, 78 RPM	7 inch, 45 RPM
Bead width	0.48 mm	0.48 mm
Bead spacing	0.528 mm	0.528 mm
Amplitude	0.10 mm	0.10 mm
Land	0.48 mm	0.48 mm
Pitch	1.688 mm	1.688 mm
Turns	47	28
Play time	36 s	37 s
Groove speed, inner / outer	327 / 980 mm/s	165 / 386 mm/s
Nozzle-smear bandwidth, inner / outer	681 / 2042 Hz	344 / 805 Hz
Stylus geometric limit, inner (3 mil tip)	684 Hz	345 Hz
slope budget at inner radius, 45°	52 mm·Hz (full amplitude legal to 520 Hz)	26 mm·Hz
G-code segments per revolution, inner / outer	2513 / 7539	2199 / 5152

A coincidence worth noting: with a 3 mil tip, mm, almost exactly one bead width, so the stylus filter and the nozzle-smear filter land at the same frequency.

Play time is the price of the lateral cut: 36 s versus v1’s 64 s on the same disc, because every pitch pays plus land. 45 RPM gives about the same time at 0.5x the bandwidth. Variable pitch and CLV both claw time back.

The encoder

The software derives the groove, writes the G-code, then decodes that same file back into audio to prove it.

• Conditioning. Mono, highpass, lowpass at the inner-radius smear limit, 6:1 compressor to 18 dB, normalize ignoring filter edge transients. The edge-transient bug silently scaled every record to 79% until the round-trip test caught it.
• Sampling. The spiral is sampled at constant arc length (0.1 mm), so outer turns get finer sampling automatically. v1’s “per-turn adaptive bandwidth” extension is free on the geometry side.
• Slope limiter. Runs on after mapping, because the bound depends on radius. Gain envelope is a running minimum over 1 mm, smoothed twice, iterated to convergence. Reports the fraction of samples touched.
• Feedrate scheduling. Printer XY acceleration bounds exactly like the stylus does, and print speed is the free knob. Each segment’s speed is set from the bead centreline curvature ( mm/s²) with 1 mm lookahead, capped at 30 mm/s. Loud passages print slowly, quiet ones fast. This is the print-time optimizer.
• Base disc. Sliced in Bambu Studio and exported as G-code. The tool reads disc top Z, centre and extrusion mode from the export and inserts the groove before the end G-code, reusing the printer’s own start sequence, bed levelling and purge.
• Self-describing G-code. Every parameter is written as ; OV key=value comments and each bead pass is labelled. The preview decodes the file that goes to the printer back into a WAV and a PNG. Tests check the decoded signal matches the internal one to under 2 µm.
• Validation gate. Exit code 2 if any check fails: bed bounds, minimum segment length, volumetric flow, planner segment rate, feedrate range, groove slope, land width, inner radius, disc fits bed, outer bead inside disc.

The four panels above are decoded from the G-code file itself, not from the encoder’s internal state. Top right is the pair of beads near mid-record with the wiggle visible. Bottom left is the modulation the stylus should see. Bottom right shows the feedrate only dipping where the melody gets loud.

The same file, played as the printer will run it. The L bead runs inward and the R bead returns outward alongside it, so the pass has no travel move. Time runs at real speed where the detail is, at the start, the turnaround at the inner radius and the end, and up to 150 times faster between.

The stress disc pushes the slope limiter and the feedrate scheduler on purpose. Each step of the ladder prints at a different speed, down to 15.5 mm/s in the steepest passage.

Run	Result
30 s synthetic melody, 10 inch, mm	41 turns, 21.8 m of groove, 4.75 g of filament, groove pass 24.4 min, slope limiter untouched, max slope 34.7°
440 Hz tone, round trip through G-code	recovered 100.5 µm vs 100 µm commanded
Stress: ladder 300 to 680 Hz at mm, 7 inch, 45 RPM	slope limiter touched 4.9% of samples, feedrate dropped to 15.5 mm/s, groove pass 8.9 min

Printer profile numbers still to tune on hardware: 30 mm/s ceiling, 2000 mm/s² acceleration budget, 500 segments/s planner assumption, extrusion width factor 1.2, 1 mm bed margin for a 254 mm disc on a 256 mm bed.

CAV or CLV

A custom turntable can slow down as the needle moves inward, like a CD. That is worth about 1.6x the audio at the same worst-case quality.

CAV, constant angular velocity, is fixed RPM like every vinyl record. Groove speed under the stylus varies 3x from outer to inner turn, so bandwidth, slope budget and print resolution per second all vary 3x, and the whole disc is designed for the worst inner turn.

CLV, constant linear velocity, slows the platter as the stylus moves inward so groove speed is constant. Every turn has the same bandwidth and slope budget, and inner turns hold as much time as outer ones. It is only possible because the turntable is custom: the controller needs stylus position, either from elapsed time after a start trigger (the file knows radius versus time) or from an arm sensor.

Decision: first records are CAV so speed control is not a variable during tracking experiments. The motor controller is designed so it can ramp.

The test ladder

One print, one question. The dominant risk in v2 is tracking, not bandwidth.

Neither v1’s analysis nor the first v2 plan addressed whether a stylus stays in a crease between two round FDM beads for 40 turns. Hence coupons first.

#	Print	Question	Judged by
1	Straight bead-pair strips at spacing 1.0 / 1.1 / 1.2 / 1.3 , 1 and 2 layers, blank	What does the crease look like, does a 3 mil tip sit in it	Macro photo of a section, needle dropped by hand, drag under a stationary cartridge
2	7 inch blank-groove disc at best spacing	Tracks 30 turns without skipping? Noise floor?	Level 0 pass/fail, recorded spectrum
3	7 inch 400 Hz tone, amplitude ladder 0.02 to 0.20 mm per turn	Where tracking breaks with amplitude: the real slope limit	Distortion or skip onset
4	7 inch tone ladder 200 to 2000 Hz at fixed amplitude	Frequency response: smear and stylus filter measured	Level per step vs preview WAV
5	Two tones, then a short melody	Pitch survives? Recognizable?	Level 1 and Level 3

Level 0 is new to the validation tiers: the stylus tracks the full disc without skipping or riding out, judged before any audio content. Prints 1 and 2 only need something that spins; 3 onward need trustworthy RPM, so they wait for the rebuilt turntable rather than an improvised spinner.

Open assumptions

• The P1S accepts a spliced plain .gcode from the SD card. Fallback: repack into a .gcode.3mf.
• Bambu Studio end-of-print markers, detected from memory, tested only against a synthetic base file.
• A 3 mil stylus in a 1.1 crease actually tracks.
• XY repeatability 0.01 mm, planner 500 segments/s, extrusion width factor 1.2.
• Bead-edge ripple spectrum, unmeasured.

Phases

Phase	Deliverable	State
0	Package, tests, docs	Done
1	Lateral emitter, splice, first-order preview	Done, unprinted
1.5	Tracking coupons and a coupon command	Next
2	Deposition heightmap, sphere stylus and cartridge model simulator; per-radius lowpass	After coupons
3	Test discs 2 to 5 through the turntable, fit the simulator	Needs turntable
4	Turntable rebuild: closed-loop motor and encoder, RPM UI with CLV-capable ramp, ceramic preamp and line out on one PCB, printed platter, tonearm, swappable headshell	In progress
5	Variable pitch, CLV records, continuous-Z experiment branch	Later

v1: what was built and played

Two v1 pipelines exist. One printed and played. The other is the design the rest of this page derives from, and its one print did not track.

The record that played came from Amanda Ghassaei’s Processing record generator, modified for a 10 inch disc at 150 RPM on the P1S with a 0.4 mm nozzle and 0.08 mm layers. The audio was a mono WAV decimated by 11 to about 4 kHz with no filter, then peak-normalized to 4-bit integers. The sketch puts 1,600 samples on each revolution and cuts a V groove 2.9 mm wide at 5.1 mm pitch, with the floor set by the sample: eleven floor heights 0.18 mm apart, 1.0 to 2.8 mm below the surface of a 3.0 mm disc. Nine revolutions fit between the edge and the 86 mm radius where the sketch stops, so the disc holds 3.6 s of the 28.7 s file. The mesh is 102,400 triangles with six open edges where the spiral starts, which the slicer’s vertex welding closed on import. It played, and the needle got stuck in it: a 3 mil tip bottoms in a V that deep.

The turntable was built by Samantha Chan (mechanical CAD) and Niegel Fernandes (electronics) for 150 RPM. The playback transfer function was never measured.

The designed pipeline is wav_to_record.py: a hill-and-dale groove on an Archimedean spiral, the floor height in 0.08 mm steps, every parameter derived from nozzle diameter and layer height, and the signal processing designed last. The first disc from it: 254 mm, 2.40 mm tall, 28 layers, 5 h 17 min, 111 g of PLA. The needle struggled to hold its spiral because the groove at the surface was almost nothing, and the sliced G-code shows why. The groove band runs from 100 to 119 mm radius with modulated floors between 1.28 and 1.92 mm, but every layer above 1.92 mm is solid across the whole disc and the top layer is a continuous skin with gap infill over the band. The mesh had quads across the top of the V, so the slicer printed the grooves as tunnels under a 0.5 mm roof and left the stylus the roof’s texture. The mesh does not survive; the slice does. The numbers below are the design at its last version, with a 0.2 mm nozzle, and were never printed.

	Inner radius (40 mm)	Outer radius (120 mm)
Steps per revolution (0.2 mm nozzle)	1,047	3,142
Sample rate	1,361 Hz	4,085 Hz
Nyquist	681 Hz	2,042 Hz

Stage	Parameter	Derivation
Highpass	300 Hz, 4th-order Butterworth	Content below 300 Hz has no harmonics inside 681 to 2042 Hz
Lowpass	681 Hz, 4th-order Butterworth	Inner-radius Nyquist as a conservative global cutoff
Compression	≥ 6:1	Fit 24 dB of input into the 18 dB the slope constraint allows
Dither	TPDF only	Zero Nyquist headroom, so noise shaping has nowhere to put the noise

The 300 to 681 Hz passband means most musical fundamentals could not be encoded directly. The ear reconstructs pitch from harmonics 2 to 4 if the signal is monophonic and periodic, so the useful input range was roughly A4 to C#5.

The mesh

The hard part of v1 was turning a waveform into a watertight, printable solid. The groove follows an Archimedean spiral . At each angular step the sample sets the floor height. The mesh must be a closed 2-manifold, every edge shared by exactly two triangles, or the slicer rejects it.

• Five vertices per step: outer and inner wall top at the surface, floor centre at , and outer and inner wall bottom at . Consecutive steps form the quad strips for the two walls and the floor.
• The land between turns is bridged top and bottom, by indexing back one full revolution at every step. Skip it and each turn is a separate shell floating in space. Ten triangles per step with the bridges.
• Four boundaries to seal: the rim to the record’s edge, the inner disk to the spindle hole, and the open notch of the V at each end of the groove, one triangle each. The rim’s last quad and the inner disk’s first quad are split where the groove starts and ends, or the seams are T-junctions.
• Sign convention. High sample values map to a deeper floor, never above the surface, so turns can never self-intersect.
• Checked, not assumed. For the 60 s record at 78 RPM with a 0.2 mm nozzle, 81,666 steps and 825,046 triangles, every edge in exactly two triangles, all facing out.

The toy record above has 48 steps per turn so the topology is visible; the construction is the same as the real one. The committed generator does not pass that check. Its land quads sit across the top of the V and roof the trench, which is what the February disc printed, and the groove spans the whole pitch so the bridges are degenerate. The animation and the numbers above use the corrected generator in anim/v1_mesh.py.

Stylus contact

Hertzian contact between an 18 µm tip and PLA at 4 g: contact radius about 5.2 µm, peak pressure about 700 MPa, roughly 14x PLA’s yield of about 50 MPa. The first play plastically sets the groove floor and effective depth resolution drops from 4 bits toward 3. Abrasive wear is a separate, much slower mechanism.

A spherical tip also cannot resolve floor features shorter than along the groove. For an 18 µm tip that limit sits about 4x above the print-limited Nyquist, so in v1 the printer, not the stylus, was the bottleneck. With a 3 mil tip in v2 the two limits coincide.

Manufacturing limits that no longer apply

• Z quantization. 0.08 mm layers gave 16 levels over 1.28 mm, with the bottom 1 to 2 bits suspect at mm Z accuracy.
• Slicer segment merging. Inner-radius steps of 0.24 mm sat at the slicer’s minimum segment length and were dropped from the G-code.
• Staircase noise. The layer staircase produced a tone at , about 4 kHz at the inner radius.
• Gap infill. The land between grooves was 0.13 mm, thinner than the nozzle, so the slicer filled it as a gap. A groove has to survive slicing, not just meshing.

None of the four is relevant once the groove is written as G-code on the XY plane.

References

• Ghassaei, A. (2012). 3D Printed Record. Objet Connex 500 resin, vertical modulation chosen because that printer was most precise in Z; about 11 kHz sample rate, 5 to 6 bits. Notes that mono records are normally cut laterally for better quality and dynamic range. Our situation is hers inverted: the P1S is most precise in XY.
• IEC 60098, analogue audio disc records and reproducing equipment.