Every synthesizer is secretly doing Fourier's job. A sawtooth wave is not a primitive building block. It is a shorthand for an infinite sum of sine waves at 1x, 2x, 3x... the fundamental, with amplitudes falling as 1/n. A square wave is the same sum restricted to odd harmonics. A triangle wave adds a phase inversion on every other partial. These shapes are efficient fictions, ways to generate complex spectra without running hundreds of oscillators simultaneously.
Additive synthesis drops the shortcut and does the job explicitly: one oscillator per partial, each with independent amplitude and frequency control. It is the most mathematically honest form of synthesis. It is also the most computationally expensive, which is why it nearly disappeared, and why its revival is instructive.
What Fourier Actually Said
In 1822, Jean-Baptiste Joseph Fourier published his proof that any periodic function, including any repeating audio waveform, can be represented as a sum of sinusoids at the fundamental frequency and its integer multiples:
f(t) = A₀ + Σ Aₙ · sin(2π · n · f₀ · t + φₙ)where f₀ is the fundamental, n is the harmonic number, Aₙ is the amplitude, and φₙ is the phase of the nth partial.
This is not just a mathematical curiosity. It is the reason why timbre exists. Two instruments playing the same note at the same volume sound different because their Aₙ values differ. The distribution of energy across harmonics is the spectral fingerprint of an instrument.
Change those amplitudes dynamically over time, and you get the attack transient, the body of the sustain, and the release tail that make a piano sound like a piano and not a sine wave approximation of one.
The Hammond Organ: Additive Synthesis in Hardware (1935)
Laurens Hammond was an engineer, not a musician. In 1935 he manufactured the first Hammond organ, built around 91 spinning tonewheel generators, each producing a sine-like tone at a specific frequency. Each key connected to a set of these tonewheels, and the player controlled the amplitude of each harmonic with nine sliding drawbars.
The nine drawbars correspond to these traditional pipe-length designations: 16', 5⅓', 8', 4', 2⅔', 2', 1⅗', 1⅓', 1'. In harmonic ratios relative to the 8' fundamental:
| Drawbar | Pipe Length | Ratio | Description |
|---|---|---|---|
| 1 | 16' | 0.5x | sub-octave |
| 2 | 5⅓' | 1.5x | quint |
| 3 | 8' | 1x | fundamental |
| 4 | 4' | 2x | 2nd harmonic |
| 5 | 2⅔' | 3x | 3rd harmonic |
| 6 | 2' | 4x | 4th harmonic |
| 7 | 1⅗' | 5x | 5th harmonic |
| 8 | 1⅓' | 6x | 6th harmonic |
| 9 | 1' | 8x | 8th harmonic |
Notice what is missing: the 7th harmonic. The 7th partial (7x the fundamental) falls between available tonewheel frequencies. No generator was close enough, so Hammond left it out entirely. The 7th harmonic is the one that gives acoustic instruments a slightly rough, untempered quality. Without it, the Hammond sits slightly cleaner than nature intended.
The 5⅓' drawbar at 1.5x is not a harmonic of the note being played. It is the perfect fifth above the fundamental, which means it sits outside the natural harmonic series for that note. Including it adds weight that Fourier's recipe for a natural instrument would not produce, contributing directly to the Hammond's character.
Each drawbar has nine positions (0 through 8). Pull all nine to maximum and you get a dense stack of partials. Weight the sub-octave and quint while pulling back the upper harmonics and you get the thick bass register that made the B-3 indispensable in soul and jazz. The players doing this in 1965 were performing real-time additive synthesis without knowing the term.
Going Digital: 126 Harmonics Per Voice
In 1987, Kawai released the K5, a digital synthesizer built around additive synthesis at a scale the Hammond could not approach. Where the Hammond had 9 drawbars per voice, the K5 offered 126 independent harmonics, each with its own six-stage envelope.
That last detail is the critical advance: not just independent amplitude per harmonic, but independent amplitude over time. On a Hammond, setting a drawbar to position 6 fixes that harmonic at a constant level for the entire note. On the K5, the 5th harmonic could start silent, peak at attack, decay to a sustain level different from the fundamental, and release at a rate entirely its own, independently of every other partial.
With 126 harmonics and 16-voice polyphony, the K5 was running roughly 2,000 simultaneous sine wave oscillators using dedicated digital hardware. The synthesizer was notoriously difficult to program because its interface presented all 126 harmonics as a bar graph you edited one by one. It never sold well. The sounds it produced were genuinely unlike anything else.
Kawai returned to additive synthesis in 1996 with the K5000. This version offered up to 64 harmonics per source, six sources per patch, and a 128-band formant filter for shaping the composite spectrum. A single K5000 patch could stack hundreds of harmonics with independent envelope control. At 32-voice polyphony, the computational scope was enormous for dedicated hardware.
The Problem With Pure Additive Synthesis
The K5 and K5000 exposed two limitations that kept additive synthesis from becoming the default approach.
CPU cost scales linearly with partial count and sample rate. Running 64 harmonics at 44,100 Hz means generating roughly 2.8 million sine wave samples per second per voice. Multiply by polyphony and the numbers quickly exceeded what software running on consumer hardware could manage through the late 1990s. This was not a problem that clever algorithmic shortcuts could fix. More harmonics genuinely means more computation.
Programming complexity is the second problem. Sculpting a convincing piano by drawing 126 harmonic envelopes manually requires knowing which harmonics attack first, which decay fastest, and where the inharmonicity is concentrated. This information is not intuitive. You can spend an afternoon editing bar graphs and produce something that sounds vaguely like a marimba.
The practical result: additive synthesis remained a specialist technique, admired more than used.
SMS: Additive Synthesis in Reverse
In 1990, Xavier Serra and Julius Smith at CCRMA (Stanford) published Spectral Modeling Synthesis, which inverted the programming problem. Instead of specifying harmonics manually, analyze an existing sound and extract the harmonic content automatically.
SMS decomposes a signal into two components:
Deterministic: A set of time-varying sinusoids, each with tracked amplitude and frequency trajectory. This is additive synthesis, with the partial parameters derived from analysis rather than drawn by hand.
Stochastic: A noise residual representing everything the sinusoids cannot capture. Breath noise, bow noise, consonant transients, any component that does not behave like a stable sinusoid ends up here, modeled as filtered noise.
The deterministic part can be manipulated independently. Stretch time without changing pitch by advancing the sinusoidal envelopes more slowly while keeping frequencies fixed. Shift pitch without changing timbre by scaling all frequencies while keeping envelope shapes intact. Morph between two sounds by interpolating their sinusoidal models.
This is the underlying mechanism in high-quality time-stretching algorithms and in some of the more capable AI audio processing tools. The connection to additive synthesis is direct: SMS is what happens when you use analysis to automate the parameter estimation that makes hand-built additive synthesis so tedious.
320 Partials: Native Instruments Razor (2011)
By 2011, general-purpose CPUs were fast enough to make large-scale additive synthesis practical in software. Native Instruments Razor, designed by Errorsmith, used 320 partials per voice. That is more than twice the K5's harmonic count and roughly five times what real-time CPU synthesis managed in the K5 era.
Razor's interface addressed the programming complexity problem differently than Kawai did: instead of editing individual harmonic envelopes, it provides effect modules that transform the entire spectrum algorithmically. A dissonance module shifts partials slightly out of tune in a controllable pattern. An overtone module multiplies amplitudes by a formula. These meta-controls shape harmonic content without requiring direct manipulation of 320 individual parameters.
The sounds Razor produces are distinctively additive: very clean in the upper register, with a precision that analog emulations cannot match, and capable of spectral textures that exist because specific partials are held static at specific amplitudes. Patches with those characteristics are not achievable in subtractive or FM synthesis without approximation.
What Only Additive Synthesis Can Do
Every other synthesis method controls the spectrum indirectly. Subtractive synthesis applies a filter to a source waveform. FM synthesis creates sidebands through carrier-modulator relationships whose spectral distribution follows Bessel functions and is hard to reason about intuitively. Wavetable synthesis cycles through pre-recorded spectra.
Additive synthesis gives direct, independent control of every partial. This makes certain things straightforward that are genuinely difficult otherwise:
Non-harmonic partial spacing: Stretch the 3rd harmonic to 3.1x the fundamental and you introduce inharmonicity. This is what makes a bell sound like a bell and not like a Fourier-perfect harmonic stack. Piano strings are also inharmonic in their upper partials, for physical reasons. Additive synthesis models this directly by detuning individual partials.
Asymmetric transient envelopes: Design the attack so that high harmonics arrive before the fundamental. This happens in bowed strings and in some brass attacks. In subtractive synthesis you are fighting the filter to achieve this. In additive synthesis you just set the envelope start delay differently per partial.
Spectral freezing and morphing: Hold a single harmonic snapshot indefinitely by setting all envelopes to sustain mode, or interpolate between two harmonic configurations over time for a morph that no waveform table can render cleanly.
Silence at specific harmonics: Remove the even harmonics and you get a hollow clarinet-like quality. The control is exact, not approximate.
The Hammond organ gave players eight amplitude levels across nine harmonics. The Kawai K5 gave six envelope stages across 126 harmonics. Razor processes 320 partials algorithmically in real time. Each step is the same Fourier principle executed at a different scale with a different interface.
The math has not changed since 1822. The question has always been: how many sine waves can you afford to run, and how do you give a musician useful control over that many independent parameters without the interface consuming more effort than the music.