What Is Audio Noise Reduction? A Clear Explanation
Audio noise reduction is the process of identifying and reducing unwanted sounds in a recording while preserving the sounds you want to keep, typically speech, music, or other intentional audio. It is one of the most widely used tools in audio post-production, and AI-based approaches have made it far more accessible in recent years.
What counts as “noise”?
In audio engineering, noise refers to any sound that was not intentionally recorded:
- Broadband noise: hiss, static, and tape noise that spreads across all frequencies
- Tonal noise: hum from electrical interference, typically at 50 Hz or 60 Hz and harmonics
- Impulsive noise: clicks, pops, and crackle from vinyl, damaged recordings, or electrical glitches
- Environmental noise: air conditioning, traffic, wind, rain, crowd noise
- Self-noise: the inherent noise floor of microphones and recording equipment
Different types of noise require different reduction techniques.
Traditional noise reduction methods
Before AI became dominant, noise reduction relied on signal processing algorithms:
Spectral subtraction
This method estimates the noise spectrum during a “silent” section of the recording and subtracts that estimated noise from the entire signal. It works reasonably well on stationary noise but often introduces a distinctive “musical noise” artifact, a wavering, underwater quality that many listeners find unpleasant.
Noise gating
A noise gate mutes the signal entirely when it falls below a set threshold. This works for intermittent noise between takes but leaves any noise that occurs simultaneously with the desired signal untouched.
Adaptive filtering
Adaptive filters estimate and remove noise continuously, adjusting in real time. They require a reference signal (a separate microphone capturing only noise) and are more commonly used in live sound applications.
How AI-based noise reduction works
Modern AI noise reduction, as used in tools like Noise Remover, takes a fundamentally different approach. Instead of estimating and subtracting noise, it learns to recognize speech directly.
A deep neural network is trained on thousands of hours of audio, both clean speech and speech mixed with various types of noise. Over time, the model learns to predict what the clean speech sounds like, even when buried under noise. At inference time, it runs your audio through this learned model and outputs a cleaned version.
This approach has several advantages:
- It handles non-stationary noise: noise that changes over time, like intermittent traffic, is no longer a problem.
- It preserves naturalness: rather than subtracting estimated noise (which can introduce artifacts), the model reconstructs speech.
- It works on diverse noise types simultaneously without user configuration.
The main limitation is computational cost. Running a large neural network requires significant processing power, which is why real-time AI denoising was initially only available on powerful servers. Optimized model designs and modern browser technology now make it feasible to run directly in your web browser.
Noise reduction vs. noise cancellation
These terms are often confused:
| Term | What it means |
|---|---|
| Noise reduction | Processing recorded audio in post to reduce noise |
| Noise cancellation | Hardware system (like ANC headphones) that generates anti-phase sound waves to cancel noise in real time |
Noise cancellation is a hardware technique used in headphones and some microphones. Noise reduction is a software process applied to recordings. They address different problems.
When should you use noise reduction?
Noise reduction is most effective when:
- Speech or music is clearly louder than the noise floor
- The noise is consistent enough to characterize (hiss, hum, fan noise)
- You want to clean up a recording that cannot be re-recorded
It is less effective when:
- The noise is louder than the desired signal
- The noise shares frequencies with the signal (e.g., crowd noise during speech)
- The original recording was made at very low gain
The right level of processing
More noise reduction is not always better. Aggressive processing can:
- Remove the high-frequency “air” that makes voices sound natural
- Create metallic or robotic artifacts on consonants
- Reduce intelligibility by over-smoothing transients
The goal is a clean recording that still sounds like a human being, not a perfectly sterile but lifeless one. Start with moderate settings and increase only as much as necessary.
Ready to hear the difference? Try Noise Remover on your own recording.