Audio DSP
Digital signal processing applied to sound reproduction: convert the signal with an ADC, reshape it arithmetically, convert it back with a DAC. In this spoke it is the correction layer — everything done to the signal to compensate for what the drivers, the box and the space do to it.
The spoke met DSP first as a tool (camilladsp) and as an aside in a primer (loudspeaker-primer-hackaday‘s “DSP shortcut”). This page holds the function set itself.
The functions
| Function | What it compensates |
|---|---|
| Equalization | uneven response from driver, box or room |
| Time alignment | drivers at unequal distances from the listener |
| crossover | band-splitting, done before amplification instead of after |
| Dynamic compression | volume swings, and excursion beyond Xmax |
| Acoustic correction | the environment — reflections, and noise |
Secondary functions the corpus has seen named: matrix mixing, headphone crossfeed, loudness compensation (audio-dsp-hardbag-guide).
Two things it replaces, and one it cannot
It replaces the passive network. An active crossover splits bands in software, so each driver gets its own amplifier channel. Slopes become configuration rather than soldered components (camilladsp ships Linkwitz-Riley 24 dB/octave presets).
It replaces some of the box. EQ can flatten a response the enclosure alignment did not deliver, which is why the primer calls it a shortcut: it lets a mediocre passive design measure well.
It cannot replace excursion or output. Boosting a band the driver physically cannot deliver moves the cone further and runs into Xmax and thermal limits sooner. DSP redistributes what the transducer already has; it does not add any. Nothing in this corpus states that limit directly, which is worth noting as an absence rather than as a source’s claim.
Implementation
Fixed-point versus floating-point arithmetic is the first split; the corpus holds no source explaining
what the choice costs. The hardware range seen so far runs from a Raspberry Pi 5 appliance with eight
analog outputs (camilladsp) down to RP2040/RP2350-class microcontrollers
(audio-dsp-hardbag-guide) — a spread of price and power for the same arithmetic. ../embedded-iot-wiki
owns those chips as silicon; what they compute is here.
Vendor capability figures — channel counts, filter-band totals, maximum delay — are marketing numbers until measured. The 85 ms delay spec on that page works out to about 29 m of path difference, roughly two orders of magnitude past any use in a vehicle.
Open
- Fixed versus floating point. What the choice costs in headroom, noise and cumulative filter error. No source.
- Does correction survive moving your head? Time alignment and EQ are computed for one listening position. In a room this is the standard sweet-spot problem; in in-vehicle-audio the position is fixed, which is unusually favourable. Nothing here measures either.
- Where correction stops helping. The excursion limit above is stated from physics, not from a source in this corpus.
Related
camilladsp · crossover · in-vehicle-audio · audio-dsp-hardbag-guide · loudspeaker-primer-hackaday · digital-filter-audibility