LinnDrum: sampled voices, articulation and the LM-1 distinction

Technical research · edition 2 · reviewed 2026-10-24

Prismal Labs native editor capture, illustrating the corresponding software project. View full-size image

What we found

The LinnDrum study corrected two misleading shortcuts: importing LM-1 specifications into LinnDrum, and replacing a documented segmented converter with arbitrary quantization. Roger Linn's account assigns the often-quoted 28 kHz nonlinear 8-bit description to the LM-1. That is not sufficient evidence for the successor's entire signal chain.

The reusable findings concern signal ordering and source identity. A converter can be modeled from its documented levels while the original waveform and dynamic output filter remain missing. Keeping those layers separate prevents a technically correct decoder from becoming an unsupported claim about a whole instrument.

Keep the instrument and articulation identities straight

Evidence: Roger Linn distinguishes the sampled LM-1 from its 1982 LinnDrum successor, which added longer sounds and crash/ride cymbals. His 28 kHz nonlinear 8-bit detail specifically describes the LM-1; it is not evidence for LinnDrum's exact conversion chain. LINN-DESIGN

The original LinnDrum instructions give bass two volume levels. Their tunable instrument list excludes bass, while closed-hat triggers cut off an open hat. LINN-OM

Kick inference: a rounded recorded-drum-style thump, damped nonharmonic body, short wooden beater and gently quantized texture. Provide LinnDrum / Studio and LinnDrum / Low tape. Adjustable pitch and continuous velocity are modern extensions. No LM-1 sample clock, ROM identity or exact converter law is claimed.

Later: record sidestick separately from full snare, closed/open/pedal hat relationships, tom/conga tuning and sampled cymbal tails. Keep LM-1 and LinnDrum source and capture identifiers separate.

Evidence: research/drums/MACHINES.md.

The segmented converter and signal ordering

LinnDrum: The service description places bass in the multiplexed group and routes it through the same type of VCF used by TOM/CGA. That filter opens for attack and closes as the sound decays. Our static filter remains a known gap; unmeasured filter timing was not guessed in this pass.

AM6070: AMD Table 3 supplies actual segmented decoder currents. With chord C=0..7 and step S=0..15, the unsigned level is 2^C (2S + 33) - 33, with sign handled separately. Normalized full scale is 8031. The implementation selects the nearest representable level with bounded work. It does not use the continuous logarithmic mu-law formula or assume telephone G.711 byte inversion. This is an authored encoder feeding the documented decoder levels, not a reconstruction of the original recording encoder.

For LinnDrum/DMX the authored source decay now precedes companding. MIDI velocity scales the converted signal; it no longer changes the source beater cutoff. Their continuous velocity and eight-voice option remain modern performance extensions. Full-scale normalization, source timbre, beater/modal ratios, gain and filter cutoffs are still authored values. A correct converter cannot make an incorrect source waveform match hardware. We do not label these corrections measured fidelity.

Evidence: research/drums/HARDWARE_ACCURACY.md.

The filter is dynamic, not just a fixed tone control

The original instructions allow tuning snare/sidestick, toms and congas; closed hat decay is adjustable. Snare has three programmed levels, while several other sounds have two. These are different from continuous MIDI response. Linn operating instructions, printed p.1, original text on mirror

Technical information places all voices behind PROM playback/Am6070 conversion. TOM/CGA have clock-controlled pitch and a dynamic four-pole CEM3320 filter; snare/sidestick share a playback circuit; hats have a separate VCA envelope. Linn technical information, DRM BOARD sections, original text on mirror

Current procedural snare, clap, shaker, rim and metal models do none of this. Refactor a finite procedural source only as a provisional way to test that signal chain. Keep source decay before companding and output accent gain after it; a velocity-dependent synthesized noise source would change a recording that should remain fixed. An arbitrary four-pole low-pass is not a calibrated dynamic filter. Exact coefficients require a legible schematic and reference measurements.

Evidence: research/drums/HARDWARE_REVIEW_V2.md.

Controls and shared playback behavior

The original instructions permit tuning of snare/sidestick, toms and congas, approximately one octave in total, and adjustment of closed-hat decay. Other voices are fixed-pitch. Its mixer has level/pan controls, and programmed trigger levels differ by voice. Operating instructions, printed p.1, original text on mirror

The existing continuous velocity exponent is an extension. A stock-level mode needs explicit trigger-level policy and evidence for its gains; distinguish that mode from a continuous velocity response curve. Do not infer layer switching or dynamic filter coefficients from the number of trigger levels.

Evidence: research/drums/MODEL_CONTROL_MIGRATION_PLAN.md.

Working through the decoder law

The documented unsigned AM6070 level law is L(C,S) = 2^C * (2S + 33) - 33, for chord C=0..7 and step S=0..15, with sign separate. Thus L(0,0)=0, L(0,15)=30, and L(7,15)=8031. Dividing by 8031 gives the declared normalized decoder levels.

An encoder that chooses the nearest level is an authored way to drive that decoder. It does not reconstruct the historical recording encoder or imply a particular byte inversion. Testing the level table independently is more informative than comparing it with a continuous logarithmic curve and calling both “8-bit companding.”

Applying the findings

Store the source waveform independently of velocity and place gain where the evidence places it. Keep the source decay before companding when reproducing the reviewed signal order. Test quiet tails especially carefully: moving the decay after a nonlinear conversion changes the level sequence and texture.

Represent discrete programmed levels and shared voice interruption explicitly if hardware behavior is the target. Preserve continuous velocity, broad polyphony and additional pitch controls as separately named software modes. This design is equally useful for other sample-based instruments with fixed recordings and nonlinear conversion.

Limits and open questions

The original LinnDrum PCM, exact multiplex timing and calibrated dynamic filter are not established by the procedural model. Technical descriptions available through manual mirrors need a legible original schematic before deriving additional coefficients. The AM6070 equation supports the converter-level correction; it does not prove a recorded hardware match.

Research illustrations

Prismal Labs native editor capture, illustrating the corresponding software project. View full-size image

Primary references

• Roger Linn: instrument museum and development history
• LinnDrum Rhythm Programmer instructions

Research provenance

The methods and results above are drawn from these retained research records. Dates and product versions inside the article identify the scope of each experiment; publication review does not mean the hardware or numerical tests were rerun.

• research/drums/MACHINES.md — SHA-256 b5a49d507c0d8c0fc3e8c39ac967ee1600192204eced1630b9e7b48bee9c7915
• research/drums/HARDWARE_ACCURACY.md — SHA-256 9d5798d139a784814d46829caefa6c7c8c23234a8ca25780c806b53826b420b1
• research/drums/HARDWARE_REVIEW_V2.md — SHA-256 4b556e1ed4bedcd8d3649d781a7c1bd8fab0693939519ce27e387c11a0e60e42
• research/drums/MODEL_CONTROL_MIGRATION_PLAN.md — SHA-256 79850a171c8ff6481addbaed4d6700c75821dc00427d755aa4dd333e09dd2af5

Related projects

KickGenerator · SnareGenerator · HiHatGenerator · TomGenerator · CymbalGenerator · ClapGenerator · RimGenerator · CongaGenerator · ShakerGenerator · TambourineGenerator

All audio and hardware research · Audio Plugins