A direct-conversion (zero-IF) receiver mixes the radio signal straight down to 0 Hz. Small imperfections then leave a constant offset in the I and Q samples, which shows up as a spike at the centre of the spectrum. Most low-cost receivers in a Software Defined Radio (SDR) work this way.

Where It Comes From

  • local oscillator leakage: a little of the oscillator signal leaks into the mixer's radio input and mixes with itself, which produces a DC term
  • mismatch in the mixers, amplifiers and ADCs
  • strong nearby signals can make the offset change over time

Why It Matters

  • the centre of the band is often where the wanted signal is
  • a large spike at 0 Hz dominates a dB scale and hides weaker detail
  • a machine learning model can learn the spike as a feature, because it is in every capture from that radio

Removing It

In hardware:

  • many receiver chips run an automatic DC correction
  • tuning offset: tune the oscillator slightly away from the signal, so the spike lands where nothing is, then shift the signal back digitally

In software, with a moving average:

  • this is a simple high-pass filter, done separately on I and Q
  • a cumulative sum computes every window average with one subtraction, so the cost does not grow with
  • subtracting each STFT frame's own mean removes whatever is left in that frame
  • if the I and Q paths have slightly different gains, or a phase difference that is not exactly 90°, a mirror image of each signal appears at the opposite frequency
  • receivers calibrate it out, and machine learning pipelines often add random IQ imbalance as a data augmentation, so a model works on many different radios