Noise reduction

A MATLAB adaptive filter designed to reduce background noise while preserving important environmental sounds.

Overview

In a team of four, we implemented a Least Mean Squares (LMS) adaptive filter in MATLAB and evaluated it on recordings from Carnegie Mellon’s campus. The goal was to reduce background noise while keeping important environmental sounds distinguishable. The project reported up to 95% lower waveform amplitude across its evaluation.

Original and processed audio waveforms with time axes and separate amplitude scales
Fig. 1 Original and processed waveforms with different amplitude scales

Algorithm development

We first explored cancellation using an inverted waveform. This illustrated destructive interference, but a fixed waveform would not track changing background noise. We also explored a recursive FFT and band-pass filtering to inspect and filter frequency content.

We instead implemented an LMS filter that updates its coefficients using an error signal. Unlike the initial fixed-waveform approach, the filter can adapt its response over time.

We varied filter length and step size to explore convergence speed, stability, and filtering performance. Plots and audio playback supported comparison of the original and processed recordings.

Illustrative LMS desired-signal and array-output curves approaching one another over iterations
Fig. 2 Desired signal and LMS output over successive iterations

Implementation & testing

We converted the recordings to single-channel signals, generated a representative noise signal, and applied the LMS filter to produce processed audio. Gym, bus, lawn, and study-area recordings tested the approach across different background sounds.

Waveform examples from four campus recordings used to evaluate filtering across different environments
Fig. 3 Original and processed waveforms from four campus recordings

We compared waveform amplitudes and listened to the processed recordings to assess whether important sounds remained distinguishable. These were visual and listening checks rather than a controlled measure of selective noise removal.

Results

The project reported up to 95% reduction in waveform amplitude, with important sounds remaining distinguishable in listening checks.

Tools

MATLAB