The Noise Reduction Problem Nobody Talks About
Search "how to remove background noise from audio" and you'll find dozens of articles recommending free plugins and AI tools that promise to "eliminate noise in one click." Some of them genuinely work well for simple, consistent noise. Most of them — used incorrectly — produce audio that sounds worse than the original recording.
This is the fundamental tension in noise reduction: every algorithm that removes noise also removes some of the signal you want to keep. The goal is to remove as much noise as possible while removing as little of the desired audio as possible. How well a tool does this depends entirely on the type of noise, the type of signal, and how the tool is used.
Types of Background Noise (They Are Not All the Same)
Most guides treat "background noise" as a single thing. It isn't. The technique that works perfectly for one type of noise will make another type worse.
Stationary broadband noise (tape hiss, HVAC hum, electrical buzz)
This is the most "treatable" category. Stationary noise has a consistent spectral profile — the same frequencies at the same level throughout the recording. Modern spectral noise reduction tools can profile this noise from a section of silence and then subtract it from the entire recording. This is what "noise reduction" in Audacity does, and it works reasonably well for this specific category.
Intermittent broadband noise (traffic, rain, crowd)
This is harder. The noise level and spectral content varies constantly. A spectral profile taken from one moment doesn't represent the noise at another moment. Tools that work by profiling a "noise floor" struggle here and tend to leave artifacts — the characteristic "underwater bubbling" or "digital warble" that makes processed audio sound unnatural.
Impulse noise (clicks, pops, crackle)
This is a completely different class of noise. Clicks and pops are extremely short-duration transients — often 0.5–5 milliseconds — that contain broadband energy. Broadband noise reduction does nothing to them. They require a dedicated declicking algorithm that identifies and replaces each impulse individually. Vinyl records, digitized tapes, and old broadcast recordings typically have this problem.
Resonant noise (AC hum at 50/60 Hz and harmonics)
Electrical interference from power supplies creates a fundamental tone at 50 Hz (Europe) or 60 Hz (North America) plus integer harmonics. This sounds like a constant low-pitched buzz and is separate from broadband hiss. It requires notch filtering at specific frequencies, not broadband noise reduction.
Reverberation ("room echo")
Technically not "noise" but often treated as such. Reverb is what you hear when a recording was made in a reflective room — a smearing and decay of every sound. It is extraordinarily difficult to remove because it is physically mixed into the signal: every reflection overlaps with subsequent direct sound. AI dereverberation tools exist and have improved dramatically, but results depend heavily on the recording.
What the "Noise Floor" Actually Means
The noise floor of a recording system is the level of noise produced by the system itself with no input signal. It's measured in dBFS (decibels relative to full scale) — a typical professional audio interface has a noise floor around -120 to -130 dBFS. A consumer laptop's built-in microphone input might be -70 to -85 dBFS. The difference is not subtle — it's a factor of 10,000 to 100,000 in noise power.
Once noise is baked into a recording, it cannot be "uncaptured." Noise reduction works by identifying which parts of the frequency spectrum are below an assumed threshold and attenuating them. The problem is that audio signals also contain energy below that threshold — quiet consonants, room tone, breath, reverberation tails. Aggressive noise reduction removes these along with the noise.
Signal-to-noise ratio (SNR) is the key metric. If a recording has an SNR of 30 dB (signal is 1,000 times more powerful than noise), noise reduction is relatively easy. If the SNR is 10 dB (signal is only 10 times more powerful), even sophisticated algorithms have to make painful tradeoffs.
Free and Affordable Tools Worth Using
Audacity Noise Reduction (Free)
Works best for stationary broadband noise. The process:
- Select a section of the recording with only noise (no speech or music) — even 0.5 seconds is enough
- Effect → Noise Reduction → Get Noise Profile
- Select all audio → Effect → Noise Reduction
- Start conservatively: Reduction 12 dB, Sensitivity 6, Smoothing 3
- Preview before applying — if you hear the "watery" artifact, reduce the Reduction amount
The most common mistake: setting reduction too high. 12–18 dB of broadband reduction is usually the practical limit before artifacts become audible.
Adobe Podcast Enhance Speech (Free, Web-Based)
Adobe's AI-powered tool works surprisingly well for speech-only recordings with background noise. It was trained on a large corpus of speech data and handles intermittent noise better than most filter-based tools. Limitations: it only works for speech (not music or mixed content), and the maximum file length is limited on the free tier.
iZotope RX Elements (~$99)
The entry-level version of iZotope's professional suite. Includes the Dialogue Isolation module for voice separation and a decent noise reduction module. This is the most capable tool in this price range and is what many podcasters and post-production editors use.
DaVinci Resolve (Free for basic version)
Includes Fairlight audio tools with a usable noise reduction module. Good if you're already editing video and don't want another application.
What You Cannot Fix with Free Tools
Severe noise + low-SNR recordings
When the noise is almost as loud as the signal, no consumer tool produces acceptable results. You need spectral editing — the ability to look at a detailed time-frequency display of the recording and manually identify and attenuate noise in specific frequency bands at specific moments, while protecting the signal.
Noise that overlaps spectrally with your audio
If you're trying to remove traffic noise from a recording of someone speaking outdoors, you have a problem: both the traffic and the voice contain substantial energy in the 200 Hz–2 kHz range. Any reduction in that range attenuates both. Professional tools handle this with harmonic extraction — identifying the harmonic structure of the voice and protecting it specifically — but this is genuinely hard to do well.
Vinyl record clicks and crackle at scale
A badly worn vinyl record can have hundreds or thousands of clicks per minute. You can manually repair individual clicks in any DAW by drawing in a correction, but doing this for an entire album by hand would take weeks. Professional declicking processes the entire recording automatically, identifies each impulse, and interpolates a replacement without the click. iZotope RX's full suite (~$400+) has this capability; the free tools don't.
Recordings with multiple overlapping noise problems
A tape that has hiss, some mold-related crackle, wow-and-flutter, and also captured room reverb requires four different processing chains applied in the right order. Get the order wrong and each step makes the next harder. This is the domain of professional restoration engineers who have processed thousands of hours of material.
The "AI Noise Removal" Reality Check
Several well-funded startups now offer AI-powered noise removal — Adobe Podcast, Krisp, NVIDIA RTX Voice, Cleanvoice, and others. They're genuinely impressive for their intended use case: real-time communication and podcast recording where the noise is speech-interfering ambient sound.
They are not the right tool for:
- Music recordings
- Archival material (tapes, vinyl, old broadcasts)
- Recordings with structural damage (dropouts, severe distortion)
- Material where audio quality genuinely matters
AI models are trained on specific distributions of noise and speech. When the input falls outside that distribution — a 1970s field recording, a warped vinyl transfer, a telephone interview from 1990 — the model's output becomes unpredictable. Sometimes it's fine. Often it introduces strange artifacts that are uniquely difficult to describe or explain to a client.
When to Use a Professional Service
The honest answer: when the recording matters and you've already tried the tools above without getting a result you'd be happy with.
A professional restoration engineer has access to tools that cost $5,000–$50,000 (Cedar Audio suites, Sonic Solutions NoNoise, full iZotope RX Advanced with manual editing capabilities) and more importantly has the trained ear and experience to know what tradeoffs to make.
The practical test is simple: listen to the audio you got after running it through your tools. Would you be okay sending that to someone? Would you be okay with that being the only version that exists?
If the answer is no, it's worth having a professional look at it.
At WefixSound, we provide a free restoration sample — a cleaned-up segment of your actual audio — before you decide whether to proceed. You hear the result before paying anything. Submit your recording here and we'll show you what's recoverable.
Summary: The Right Tool for the Right Problem
| Noise Type | Free Tool | Paid Tool | Professional |
|---|---|---|---|
| Tape hiss (stationary) | Audacity | iZotope RX Elements | Severe cases |
| HVAC/electrical hum | Audacity EQ notch | iZotope RX | With other issues |
| Vinyl clicks/crackle | Manual (slow) | iZotope RX | Large collections |
| Traffic/crowd (variable) | Adobe Podcast (speech only) | iZotope RX | Complex material |
| Wow and flutter | No | iZotope RX Advanced | Most cases |
| Reverb removal | No | iZotope RX | Severe cases |
| Multiple overlapping issues | No | Difficult | Yes |
The best noise reduction is the kind you never had to do — which means: record in the quietest environment you can find, as close to the source as possible, with the best input chain available. But for everything that already exists and can't be re-recorded, the tools above are where to start.