Frequency-Domain Effects

Effect: Tremolo / Vibrato

  • Vibrato: Oscillate the pitch of a signal. Resample as an effect (later), or just mess with the instrument

  • Tremolo: Oscillate the volume of a signal. Easy effect

  • But…: "The tremolo arm on your favorite guitar, for example, is actually a vibrato arm." (source) So these terms are used interchangeably.

Effect: Wah

  • Classic frequency-domain effect

  • Idea: lowpass filter with variable passband

  • Human vocal tract is subtractive via low-pass filters: "ooh" is low-pass, "aah" is less low-pass (note inverse connection to speaker size)

  • My sample implementation uses FIR filters

  • More typical implementation is with an IIR filter cascade (need to keep number of coefficients small to be able to update the filter in reasonable time). Biquad sections are a popular choice, as here.

  • See above for fancy demos

Effect: Flanger / Phaser / Chorus

  • Flanger: Mix original signal with varying delay of few ms; creates varying "comb filter" by phase cancellation (Nyquist plugin)

  • Phaser: Simulation of flanger using cascade of biquad "all-pass filters" to get multiple phase cancellations. Originally an analog flanger substitute, but different enough to survive (Audacity effect)

  • Chorus: Use a multi-tap delay line and vary the tap positions to get varying delay and frequency shift. Less pronounced than ensemble effect (Nyquist plugin)

  • These terms are used pretty interchangeably; no standard vocab

Effect: Resampling

  • Not normally an "effect": just a thing. We've talked about it in detail.

  • As an effect: "sped up" (Chipmunks) or "slowed down" (sleepy) version of the sound

Effect: Frequency Stretching

  • Frequency domain equivalent of resampling

  • Idea: lengthen or shorten a signal while keeping its harmonic content the same

    • Time domain: Try to find the period of the signal, interpolate whole periods; useful period may be impractically long

    • Frequency domain: Take DFT, scale it up or down (interpolate / decimate in frequency domain), take inverse DFT; Fourier uncertainty is a problem, multiscale is hard

  • The difficulty of this plan is why video is time-scaled rather than audio

  • AKA pitch-shifting, because pitch is log-frequency, so multiplication becomes addition

Last modified: Tuesday, 28 April 2020, 1:28 PM