Blog / Sound Design

The State Variable Filter: Why the Oberheim SEM Doesn't Sound Like a Moog

The State Variable Filter: Why the Oberheim SEM Doesn't Sound Like a Moog

In 1974, Tom Oberheim released the SEM, a small synthesizer expansion module built around a two-pole state variable filter. Around the same time, Bob Moog's transistor ladder filter was establishing itself as the dominant voice in synthesis. Both could sweep a cutoff, both could scream into self-oscillation. But they do not sound the same, and the reason is topological, not aesthetic.

Two Integrators in a Loop

A Moog ladder filter is a cascade of four single-pole RC lowpass sections connected in series with a feedback path. The circuit is a single functional unit. Change the resonance and it interacts slightly with the cutoff. Push it into self-oscillation and you get a sine wave with character: harmonically colored because the transistors in the ladder compress at high signal levels.

A state variable filter works differently. It consists of two integrators connected in a feedback loop with a summing node. The key consequence: the output of each integrator and the summing node are simultaneously available as independent filter outputs. From one circuit you get:

  • Lowpass from the second integrator
  • Bandpass from the first integrator
  • Highpass from the summing node
  • Notch by summing the lowpass and highpass together

Four filter modes, one circuit, running simultaneously. Cutoff and resonance are mathematically independent with no cross-interaction.

This simultaneous output is the architectural difference that most filter discussions skip. On a Moog you get one output. On an SVF you choose which signal to listen to, or blend between them. You can cross-fade from LP to BP by mixing them. You can route the bandpass to one VCA and the lowpass to another and modulate them independently. You can create a notch that tracks the cutoff by summing LP and HP at the same level.

A Brief History of SVF Hardware

The ARP 1047 Multimode Filter/Resonator, introduced in 1970 as part of the ARP 2500 modular system, was one of the first commercially available voltage-controllable multimode filters. It exposed all four simultaneous outputs on separate jacks, with resonance adjustable from 0 to a ratio described as 512:1 in bandwidth, meaning the bandpass could narrow to roughly 1/32 of a semitone. This made it useful not just as a timbral tool but as a precision resonator for acoustically modeling spaces and objects.

The Oberheim SEM (1974) is the most famous SVF in synthesis history. It offers a continuously sweepable output that morphs from lowpass through notch to highpass, taking advantage of the simultaneous outputs through a single blend knob rather than separate jacks. The SEM's filter became so associated with a particular sound, open and airy at high resonance, that Arturia's iSEM and Cherry Audio's SEM plugin exist specifically to recreate it.

In 1978, Chris Huggett designed the EDP Wasp around a distinctly unusual SVF implementation. To keep the cost under £200, Huggett used CD4069UB CMOS inverters as operational amplifiers in place of the specialized op-amps used in other designs. These CMOS inverters introduce their own nonlinearities, giving the Wasp filter a character quite different from the cleaner Oberheim circuit. Doepfer released a Eurorack version of the Wasp filter.

Serge Tcherepnin's Serge modular system, and the ARP 1047 before it, demonstrated that the SVF's multiple outputs weren't just nice to have: they changed the compositional vocabulary available. You could route a patch so the bandpass got envelope-following tremolo while the lowpass went dry to the output, creating movement that a single-output filter couldn't produce.

Why 12dB/oct Sounds Different from 24dB/oct

The Oberheim SEM operates at 12dB per octave, a two-pole rolloff. The Moog ladder is four-pole, 24dB per octave. The practical effect is not just loudness past the cutoff. At 24dB, content past the cutoff drops off sharply enough that the resonant peak tends to dominate the filtered sound. At 12dB, more of the original harmonic content bleeds through, coloring the resonance. This is why SVF-filtered sounds retain more texture and why brass and string patches often track acoustic resonances more naturally with a 12dB slope.

The SVF self-oscillates cleanly. Without the cascade of saturating transistors that characterize the Moog ladder, an SVF pushed past its resonance threshold produces a near-pure sine wave. u-he's Diva plugin, whose "Uhbie" mode models the SEM's SVF topology, is capable of clean self-oscillation across the audible range. This makes the SVF useful as a tone generator: apply a sharp pitch envelope to the cutoff CV, enable self-oscillation, and you get a pitched ping without a dedicated oscillator.

The Digital Translation Problem

When Hal Chamberlin described a digital state variable filter in his book "Musical Applications of Microprocessors" (1985), the implementation looked deceptively simple. Replace the two integrators with accumulators, replace the summing node with addition, and you have the filter in about eight lines of code. Call it the Chamberlin SVF.

The problem is stability. The Chamberlin filter is only guaranteed stable when:

F1 + 1/Q < 2

where F1 = 2 sin(π Fc / Fs). The practical ceiling for typical Q settings is roughly fs/6, about 7350 Hz at a 44100 Hz sample rate. Push past that, and depending on Q, the filter can diverge. Not subtly, not with a pleasant harmonic saturation: the output blows up to maximum amplitude. Early digital SVF implementations in hardware and software discovered this the hard way, often in the form of a sudden noise burst when a performer swept the cutoff too high during a show.

The instability occurs because Chamberlin's formulation uses a forward-Euler approximation of the analog integrators. Forward-Euler discretization is simple but maps the stable analog poles to locations outside the unit circle in the z-plane at high frequencies.

The Trapezoidal Fix

Andrew Simper of Cytomic published a solution that became widely adopted in the plugin development community. Instead of forward-Euler, Simper applied trapezoidal integration to each analog integrator, a discretization method equivalent to the bilinear transform applied locally. The result is an SVF implementation that is unconditionally stable: any cutoff frequency, any Q value, at any sample rate.

The math requires solving a small 2x2 linear system per sample rather than the sequential accumulation Chamberlin used. On modern hardware this is a negligible difference. Simper documented the derivation in "SvfLinearTrapOptimised2.pdf" on the Cytomic website, and the paper has been cited in dozens of open-source plugin implementations. It is the reason modern SVF-modeled plugins do not blow up at high resonance settings.

The paper also demonstrates how to add nonlinear saturation to the trapezoidal SVF while preserving stability, which is how modern analog-modeled plugins introduce the warmth that the clean linear SVF lacks.

Practical Implications

When choosing a filter type for a patch, the SVF versus ladder decision is a parameter with audible consequences:

Use an SVF when: you want the resonance peak to sit within a harmonically rich sound, you need multiple simultaneous outputs for independent routing, you want clean self-oscillation as a tone source, or you're modeling a natural acoustic resonance where 12dB/oct slopes are more physically accurate.

Use a ladder filter when: you need a sharp frequency wall, you want saturating resonance that compresses at high levels, or you're going for the specific character of Moog-derived timbres.

The multiple simultaneous outputs remain underused in DAW-based production because most plugin filter designs expose only one mode at a time. But in modular synthesis and in some multi-output filter plugins, the ability to send the LP to the reverb bus and the BP to the delay creates movement that a single-pole approximation can't touch.

The state variable filter is not a "softer Moog." It is a different machine built on a different mathematical foundation, with different strengths, different failure modes, and a specific historical lineage that runs from the ARP 2500 in 1970 through the SEM and the Wasp to the trapezoidal implementations running inside your plugins today.