Independent Component Analysis (ICA)

By Prof. Seungchul Lee
Industrial AI Lab at POSTECH

Table of Contents

1. The Cocktail Party Problem

You're at a crowded party. The music is loud, people are laughing, and a dozen different conversations are happening all around you. However, despite the hubbub, you're able to focus on the one voice you want to hear.

(No reason to shout. Though background noise can be distracting, the brain has the remarkable ability to track conversation and scale down unwanted noise. Courtesy of the National Archives)

Consider two conversations in a room that are happening simultaneously. How is it that the two different acoustic signals of converstaion one and two can be separated out?

$$ \begin{align*} x_1 &= a_{11}s_1 + a_{12}s_2\\ x_2 &= a_{21}s_1 + a_{22}s_2 \end{align*} $$
In [1]:
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline

import IPython.display as ipd
import librosa.display
In [2]: