The Room Answers Before the Ear Does
The Room Answers Before the Ear Does
A plucked string does not simply vibrate at one frequency and stop; it carries a whole family of harmonics, each decaying and beating against the others, and the instrument's body reshapes that family again before any of it reaches air. Physical modeling synthesis takes this seriously instead of faking it with recorded samples or additive sine banks: it builds a discrete model of the physical process that makes the sound and runs that model forward in time.
The workhorse technique is the digital waveguide. A vibrating string supports traveling waves moving in both directions at once, and at any point the total displacement is just the sum of a rightward-traveling wave and a leftward-traveling one. Because the wave equation for an ideal lossless string is linear and the wave speed is constant, a traveling wave shape simply shifts along the string unchanged as time advances — so instead of solving a partial differential equation at every point, a digital waveguide stores the traveling wave as a delay line: a circular buffer that a sample walks through, one step per time tick, standing in for one direction of travel. Two delay lines, one for each direction, meet at the ends of the string, where boundary conditions — a rigid bridge, a free end, a finger stopping the string short — determine how much of the wave reflects, and with what sign and what filtering. A rigid termination reflects a wave inverted; the reflection coefficient becomes a full digital filter when the termination is not perfectly rigid, so a wood bridge and a rigid nut behave differently at the two delay line ends. Losses along the string — internal friction, air drag — are lumped into a single low-pass filter placed once per round trip rather than distributed continuously, because a filter's effect is the same whether it acts a little bit everywhere or all at once at one point, as long as the total attenuation per round trip matches. A bowed string, a blown pipe, a struck drumhead each becomes its own network of waveguides and junctions, with the excitation — bow-string stick-slip friction, a reed's pressure-controlled valve, a mallet's nonlinear contact — injected as the nonlinear element that keeps the otherwise linear, energy-conserving waveguide oscillating instead of ringing down to silence.
None of this reaches a listener without a room, and a room does its own signal processing before the ear ever gets a chance. Psychoacoustic masking is the fact that the ear does not report everything a room's air pressure actually contains — it reports only what a busy auditory system judged worth reporting, and that judgment can be predicted and even exploited. A loud tone at one frequency raises the threshold of audibility for nearby frequencies for some tens of milliseconds after it stops (temporal masking) and for a band around it while it plays (simultaneous or frequency masking) — the basilar membrane's traveling wave response to a strong tone physically overlaps and suppresses its neural response to a weaker, nearby one, so the weaker tone is present in the air and absent from perception. Lossy audio codecs — MP3, AAC, Opus at low bitrates — are built almost entirely on this fact: a psychoacoustic model estimates, for every short time frame, the masking threshold curve across frequency, and any quantization noise introduced by compression is deliberately shaped to hide just under that curve, spending bits where the ear could actually notice and none where it structurally cannot.
Spatial hearing adds one more layer, and this is where ambisonics enters. A single microphone captures pressure at a point — no direction. To capture and later reproduce a full 3D sound field, ambisonics encodes sound not as channels tied to specific speaker positions but as a spherical harmonic decomposition of the pressure field around a point: the zeroth-order component (W) is the omnidirectional pressure itself, first-order adds three figure-eight components (X, Y, Z) capturing the pressure gradient along each axis — together, first-order B-format is the first four spherical harmonics, and higher orders add more directional resolution at the cost of more channels. Because this representation is speaker-layout-independent, it can be rotated losslessly (turn the listener's head, rotate the harmonics), and it can be decoded to any speaker arrangement or to a binaural headphone pair by a decode matrix computed for that specific layout — the recording or synthesis stage never needs to know in advance where the sound will be played back, which is exactly the same separation of concerns as a waveguide model not needing to know in advance what room it will be placed in.
Three different problems, one answer in each case: represent the wave compactly instead of simulating it point by point (the waveguide), predict what the listener will actually notice instead of reproducing everything faithfully (masking), and separate the geometry of the sound field from the geometry of the speakers reproducing it (ambisonics). In each, the room, or the ear, or the reproduction system answers a question about the sound before the raw physics is ever fully computed — because the full computation was never the point; predicting the listener was.