GitHub - ggeorgovassilis/ICA

2 min read Original article ↗

Independent Component Analysis (ICA) demo

This project demonstrates how Independent Component Analysis ICA works using Python and scikit-learn. Companion blog post for this repository: https://blog.georgovassilis.com/2025/08/20/ica-demo/

Synthetic signals demo

Mixes a composition of three sine waves with a pulse signal and then adds some noise. Two versions with different amplitudes and delays are mixed for the two receivers. Last, the two recordings are fed into ICA which reconstructs the original signals.

Simple signals

Radio recordings demo

Mixes and separates real radio broadcasts. There are three clips which were mixed together three times, one for each receiver.

  • receiver1.wav: all three WAVs mixed equally
  • receiver2.wav: no changes were made to cowbells, horse races were delayed by 20ms and the volume was reduced by -10dB, interview delayed by 40ms and volume reduced by -20dB
  • receiver3.wav: cowbells were delayed by 40ms and volume reduced by -20dB, horse races was delayed by 20ms and the volume was reduced by -10dB, no changes to the interview clip.

Real signals

What is ICA?

ICA is an algorithm that isolates signals from signal mix, as long as you have at least as many mixed signals as sources. For example, two antennas at different locations pick up two radio stations broadcasting on the same frequency. Each antenna receives a different mix. ICA can separate the original broadcasts.

Running the demos

  1. Create and activate a virtual environment: 
    python3 -m venv venv
source venv/bin/activate
  2. Install dependencies: 
    pip install -r requirements.txt
  3. Run a demo: 
    python ica_demo_simple_signals.py
# or
python ica_demo_real_signals.py

Both demos save their charts as PNGs in the same folder. The real signals demo saves the reconstructed signal in audio/reconstructed.

Results and discussion

Synthetic signal

The composite sine wave of the synthetic signal is reconstructed fairly well. The pulse signal is reconstructed in principle, the base frequency and phase are correct, but the pulse is slightly deformed.

Radio recordings

Overall, ICA isolates the source signals well. While it's not visible in the chart, the reconstructed audio clips show that long streaks of silence are filled with sound from the other signals, eg. the first two seconds of reconstructed_source_3.wav

All reconstructed recordings can be listened to in audio/reconstructed

Licenses and Credits