You recorded a video and the audio has a persistent hum, hiss, or drone underneath the speech. Maybe the air conditioning was running, a computer fan was blowing, or the microphone preamp introduced noise. The speech is intelligible, but the background noise is distracting and unprofessional.
Post-recording noise reduction can help. It works by analyzing the characteristics of the noise and subtracting them from the audio signal. The result is cleaner speech with reduced or eliminated background noise.
How spectral noise reduction works
Audio noise reduction algorithms operate in the frequency domain. They break the audio signal into thousands of frequency bands using a Fast Fourier Transform (FFT), analyze which bands contain noise versus desired audio, and reduce the level of the noise-dominated bands.
The process has two phases:
1. Noise profile learning: The algorithm analyzes a segment of audio that contains only noise (no speech or desired sound). This could be a pause between sentences where only the background noise is audible. From this sample, it builds a spectral profile of the noise: how loud the noise is at each frequency.
2. Noise subtraction: The algorithm processes the entire audio file, comparing each moment's frequency content to the learned noise profile. Frequencies that match the noise profile are reduced or removed. Frequencies that differ from the noise profile (speech, music) are preserved.
What noise reduction can fix
Stationary noise (constant, unchanging character): - Air conditioning hum: Typically a strong fundamental around 60 Hz (in North America) or 50 Hz (in Europe) with harmonics. Very responsive to noise reduction. - Computer fan noise: Broadband noise centered in the 200-2000 Hz range. Responds well to reduction. - Microphone self-noise (hiss): High-frequency noise above 4 kHz. Responds well to reduction. - Electrical hum (ground loop): Typically 60 or 120 Hz buzz. Responds very well to notch filtering. - Fluorescent light buzz: 120 or 100 Hz buzz with harmonics. Responds well.
What noise reduction cannot fix
Non-stationary noise (changing, intermittent): - Other people talking: Cannot be separated from the primary speaker reliably. - Dogs barking, doors slamming, car horns: Brief loud sounds that noise reduction either leaves in or creates audible artifacts trying to remove. - Wind noise: Broadband low-frequency noise that overlaps with speech frequencies. Removing it also damages speech. - Clipping distortion: Permanent data loss. No algorithm can reconstruct the original waveform. - Echo and reverb: Reflected sound is correlated with the original signal. Noise reduction designed for uncorrelated noise is ineffective.
The artifacts tradeoff
Aggressive noise reduction introduces artifacts: metallic, watery, or "underwater" sounding speech. This happens because the algorithm removes too much frequency content, leaving gaps in the audio spectrum that the brain perceives as unnatural.
The goal is to reduce noise enough that it becomes unobtrusive without introducing artifacts that are more distracting than the original noise. A light touch usually produces the best results. Removing 80% of the noise while keeping natural-sounding speech is better than removing 100% of the noise and producing robotic-sounding speech. The MediaBrew audio denoise tool lets you preview the result before downloading, so you can dial in the right level.