Before an acoustic signal hits the FFT, it typically passes 2–3 filter stages – anti-aliasing, drift high-pass, bandpass on the band of interest. Filters are invisible helpers without which no clean evaluation succeeds.
Two families: IIR and FIR
IIR (Infinite Impulse Response): classical analog filters, digitally emulated. Very efficient – few coefficients for steep slopes. Drawback: nonlinear phase.
FIR (Finite Impulse Response): no feedback, only forward coefficients. Linear phase (no signal distortion). Drawback: many more coefficients for the same slope.
Comparison
| Criterion | IIR | FIR |
|---|---|---|
| Compute for same slope | low (10–30×) | high |
| Phase | nonlinear | linear |
| Stability | critical (poles) | always stable |
| Implementation | medium | easy |
| Ideal use | real-time level metering | modal analysis, phase-critical work |
Standard filter types
Butterworth: maximally flat passband. Classic for clean level measurement.
Chebyshev I: steeper than Butterworth, ripple in passband.
Chebyshev II: ripple in stopband, smooth passband.
Bessel: optimal phase linearity for IIR, gentler slopes. Standard when pulse shapes must be preserved.
Elliptic (Cauer): steepest possible slopes, ripple in both bands.
Use cases
- Anti-aliasing low-pass before ADC, typ. 8th-order Butterworth or Elliptic.
- High-pass against drift on accelerometers (5–20 Hz).
- Bandpass on the relevant range to save data and CPU.
- A-weighting filter (IEC 61672) for dB(A) level metering.
- Octave / third-octave filters for noise standards.
SonicTC.DSP filter chain example
Typical chain for tool-wear monitoring:
- Analog anti-aliasing low-pass before ADC.
- IIR Butterworth 4th order, fc = 100 Hz (drift removal).
- IIR Butterworth 6th order, 1–8 kHz bandpass (tool engagement).
- FFT with Hann window.
- A-weighting (FIR) if dB(A) needed.
What to remember
- IIR efficient, FIR clean – choose by application.
- Butterworth is the safe default; Elliptic for extreme slopes.
- Phase distortion can heavily alter signals – use FIR or Bessel for modal analysis.
- Filters save data, CPU and prevent aliasing – they are not optional.