What Is Audio Noise Reduction and How Does It Work?
If you've ever recorded audio with an annoying background hum, the whirr of air conditioning, or the persistent hiss of tape noise, you've probably wondered: can this be removed? And if so, how?
Audio noise reduction is the technology that answers that question. It's one of the most powerful and widely used tools in audio post-production — and understanding how it works helps you use it more effectively and set realistic expectations for what's possible.
What Is Audio Noise?
In audio, "noise" is any unwanted sound that appears in a recording. It's distinct from the signal — the sounds you deliberately captured (voice, music, environment). Common sources:
- Thermal noise (Johnson noise): Random electrical noise generated by electronic components at room temperature. Present in all microphones and recording equipment to some degree.
- Environmental noise: AC systems, traffic, crowd ambience, machinery
- Electromagnetic interference: Power line hum (50/60Hz), radio frequency pickup
- Mechanical noise: Fan vibration, cable handling, building HVAC
- Medium noise: Tape hiss in analog recordings, vinyl surface noise
All of these create a "noise floor" — the baseline level of noise present even when the wanted signal isn't. The signal-to-noise ratio (SNR) describes the relationship between your wanted signal and this noise floor. Higher SNR = cleaner recording.
How Noise Reduction Works: The Three Main Approaches
1. Spectral Subtraction (Profile-Based Noise Reduction)
This is the classic approach used in Audacity and many older tools.
The process:
- You identify a section of "pure noise" — a moment where only the unwanted background is present, with no voice or music
- The algorithm analyzes this sample to build a "noise profile" — a map of the noise across all frequency bands
- This profile is then subtracted from the entire recording — wherever the recording's energy matches the noise profile, it's attenuated
Why it works: Most steady-state background noise (AC hum, tape hiss) has a consistent character. The noise profile captured in a quiet moment accurately represents the noise character throughout the recording.
Why it has limits: If the noise varies (traffic that changes, intermittent sounds), the static profile doesn't adapt. Overly aggressive subtraction removes not just noise but also the quiet portions of the wanted signal that happen to fall in the same frequency range as the noise — creating the characteristic "underwater" or "warbling" artifact.
2. Adaptive (Dynamic) Noise Reduction
More sophisticated than profile-based tools. Instead of using a fixed noise profile, adaptive noise reduction continuously analyzes the recording and updates its model of what the noise floor looks like at any given moment.
The process:
The algorithm assumes that periods of low signal energy contain primarily noise. It monitors the signal's energy level and, during low-energy moments (pauses in speech, quiet musical passages), re-analyzes the noise floor. The reduction adapts continuously.
Why it's better: Handles variable noise that profile-based tools miss. Works well on recordings where the background noise changes — traffic that ebbs and flows, a room where HVAC cycles.
Tools using adaptive approaches: iZotope RX De-noise (adaptive mode), Adobe Audition Adaptive Noise Reduction.
3. AI/Machine Learning Based Noise Reduction
The newest generation of tools. Instead of analyzing the noise profile, these systems were trained on vast datasets of clean speech and various types of noise. They learn to identify what "speech" sounds like versus what "noise" sounds like, and separate them.
Why it's more powerful: Doesn't need a noise sample. Can separate speech from noise even when they overlap in frequency — something that spectral subtraction fundamentally can't do. Can work on recordings where there's no clean noise sample available.
Tools: iZotope RX Dialogue Isolation, Adobe Podcast Enhance, NVIDIA RTX Voice, Waves Clarity Vx, DaVinci Resolve Voice Isolation.
Limitations: Optimized for voice. Can affect musical content. Sometimes creates its own artifacts ("smoothed" or "plasticky" sound when pushed hard). Not all AI tools handle all noise types equally well.
The Core Trade-off: Noise vs. Artifacts
All noise reduction involves a fundamental trade-off: reduce more noise → introduce more artifacts.
This is because noise and the wanted signal often share frequency content. A voice and air conditioning noise both contain energy in the 1–4kHz range. The algorithm must decide how aggressively to attenuate shared frequencies — reduce too little and the noise remains audible; reduce too much and you're also attenuating the voice.
Common artifacts:
- Watery / burbling: Fluctuating suppression as the algorithm tries to track a changing noise floor
- Metallic / tinny: Over-subtraction of frequency bands, leaving only certain harmonics
- Underwater effect: Extreme over-processing that removes too much signal
- Musical noise / tonal remnants: Isolated frequency peaks left over from imperfect subtraction
Professional engineers know where the threshold is for each specific recording. They apply just enough reduction to make the noise inaudible without pushing into artifact territory.
Types of Noise and How Well They Can Be Reduced
| Noise Type | Fixability | Best Tool |
|---|---|---|
| Steady AC/fan hum | Excellent | Any profile-based or adaptive tool |
| Electrical hum (50/60Hz) | Excellent | De-hum / notch filter |
| Tape hiss | Very good | Adaptive de-noise |
| Vinyl crackle | Very good | De-click |
| Traffic noise (steady) | Good | Adaptive de-noise |
| Variable crowd noise | Moderate | Adaptive / AI separation |
| Internet compression artifacts | Poor | AI-based tools partially help |
| Severe clipping distortion | Very poor | Cannot be fully addressed |
| Missing signal (dropout) | Varies | Short gaps interpolatable; long gaps not |
The Signal-to-Noise Ratio Limit
There's a physical limit to noise reduction effectiveness: the signal-to-noise ratio.
If a recording has a loud, clear voice against a quiet hum (high SNR), noise reduction achieves excellent results — the voice signal is strong enough that the algorithm can distinguish it clearly from the noise.
If a recording has a quiet voice nearly buried under loud noise (low SNR), even the best tools struggle. When the noise level is close to or exceeds the signal level, the algorithm can't reliably separate them. At some point, the recording becomes too noisy to be meaningfully improved — the signal isn't there to recover.
This is why some recordings genuinely can't be fixed: the signal was never captured cleanly in the first place.
Where Noise Reduction Is Applied
Standalone Applications
iZotope RX, Audacity, Adobe Audition — process audio files outside of a video editor.
DAW Plugins
Noise reduction applied as plugins on tracks in a DAW (Pro Tools, Logic Pro, Reaper) — processes in real-time during playback or in the export.
Video Editor Integration
Adobe Premiere, DaVinci Resolve, Final Cut Pro all have built-in noise reduction that works on audio within the video editing environment.
Real-Time Processing
NVIDIA RTX Voice, Krisp, and similar tools apply noise reduction in real-time during live calls and streaming. Also effective when applied to recorded audio after the fact.
How Much Can You Expect?
Realistic noise reduction outcomes:
Best case (consistent, low-level noise): Background noise becomes inaudible. The listener may not even realize the recording was noisy.
Typical case (moderate noise): Background noise is reduced by 10–15dB and becomes much less distracting, though a careful listener may still notice it. The recording is significantly more comfortable to listen to.
Difficult case (heavy noise): Noise is reduced but not eliminated. Processing artifacts may be present. The result is better than the original but not clean. Professional tools achieve better results than consumer tools in this scenario.
Extreme case (noise exceeds signal): Meaningful improvement isn't possible. The recording can be made to sound different but not genuinely cleaner.
Noise Reduction as Part of a Complete Process
Noise reduction is one tool in a broader audio restoration process:
- De-hum: Remove tonal interference first (before broadband noise reduction)
- De-click / De-crackle: Remove impulsive noise (before broadband reduction)
- Broadband noise reduction: Address the remaining steady-state noise floor
- De-reverb: Address room echo (before EQ)
- EQ: Frequency correction
- Compression and normalization: Level management
Applying them in this order — and not skipping steps — produces better results than any single tool applied in isolation.
For recordings where the noise is beyond what standard tools handle, professional audio restoration engineers bring expertise and better tools to bear. WefixSound offers a free sample before payment — hear what's possible with your specific recording before committing.
Related articles: How to Denoise Audio · How to Remove Background Noise from Audio · How to Remove Hum from Audio