Audio ProductionPodcastApril 7, 2026· 9 min read

Noise Reduction Techniques Explained: How to Remove Background Noise from Audio

Background noise is one of the most common problems in audio recording — room hum, HVAC noise, electrical hiss, fan noise. This guide explains the signal processing techniques behind noise reduction, how to get the best results, and when noise reduction helps vs. hurts your audio.

What Is Background Noise in Audio?

In audio engineering, noise refers to any unwanted signal mixed in with your recorded content. Unlike digital artifacts (clipping, encoding errors), noise is typically an analog signal picked up by the microphone alongside your intended audio.

Noise falls into two broad categories:

Stationary noise

Consistent, time-invariant noise with a stable frequency profile. Examples: HVAC hum, fan noise, electrical hiss, tape hiss, ground loops. This type is easy to remove.

Non-stationary noise

Noise that changes over time: traffic, crowd noise, a phone ringing, a door slamming. This is much harder to remove without affecting the desired signal.

Most noise reduction algorithms are designed for stationary noise. They estimate the noise's frequency profile from a silent section of the recording, then subtract it from the entire file.

Spectral Subtraction: The Classic Algorithm

Spectral subtraction (Boll, 1979) is the foundational technique used in virtually every noise reducer, from Audacity to iZotope RX. The core idea is simple: if you know what the noise sounds like, you can subtract it from the signal.

Step 1: Noise profiling

Analyze a "noise-only" section of the recording (where no speech or music is present). Compute the average magnitude spectrum of this section — this is the noise profile N(k) for each frequency bin k.

Step 2: STFT analysis

Apply the Short-Time Fourier Transform (STFT) to the entire audio signal, dividing it into overlapping frames and transforming each to the frequency domain. This produces the mixed spectrum X(k) for each frame.

Step 3: Spectral subtraction

For each bin in each frame, subtract the noise magnitude: |S(k)| = max(|X(k)| − α·|N(k)|, β·|X(k)|). α is the oversubtraction factor (typically 1.0–2.0). β is the spectral floor (typically 0.01–0.05) — prevents the spectrum from going to zero, which causes musical noise.

Step 4: Phase preservation

The original signal phase is kept unchanged. Only the magnitude is modified. This is critical — modifying the phase would create severe artifacts.

Step 5: ISTFT reconstruction

Apply the inverse FFT to reconstruct each frame with the cleaned magnitude and original phase. Overlap-add the frames to produce the full-length denoised audio.

The core equation:

|S(k)| = max( |X(k)| − α·|N(k)|, β·|X(k)| )

Where X(k) = noisy signal, N(k) = noise estimate, α = oversubtraction factor, β = spectral floor

Wiener Filtering: A Smoother Approach

The Wiener filter is a more sophisticated noise reduction technique that computes an optimal gain for each frequency bin based on the estimated signal-to-noise ratio (SNR). Instead of hard subtraction, it applies a smooth attenuation that varies per bin.

Wiener gain formula:

G(k) = SNR(k) / (1 + SNR(k))

where SNR(k) = (|X(k)|² − |N(k)|²) / |N(k)|²

This means bins with high signal-to-noise ratio (loud signal, quiet noise) get a gain close to 1 (preserved). Bins dominated by noise get a gain close to 0 (attenuated). The result is a much smoother, more natural-sounding output compared to raw spectral subtraction.

Wiener filtering produces fewer musical noise artifacts because the gain function is smooth and continuous, unlike the hard threshold of spectral subtraction. It is the preferred method when audio quality is paramount.

Musical Noise: The Main Side Effect

The most common artifact of noise reduction is "musical noise" — a metallic, tonal, warbling artifact that sounds like random musical tones appearing and disappearing in the background. It is called musical noise because isolated spectral bins sound like brief tones when they appear and vanish.

Musical noise is caused by:

  • Setting the oversubtraction factor α too high — this over-suppresses bins, leaving only the loudest ones in each frame
  • Using a hard threshold instead of a smooth mask — bins are either fully on or fully off, creating sharp transitions
  • Inaccurate noise estimation — if the noise profile doesn't match the actual noise, bins will be incorrectly attenuated

Use spectral floor

The β parameter prevents bins from going to zero. A floor of 0.01–0.05 prevents the extreme attenuation that creates tonal artifacts.

Smooth the mask

Apply temporal smoothing to the gain function across frames. This prevents sudden per-bin changes that create the warbling effect.

Lower oversubtraction

Reduce α to 1.0–1.2. More aggressive noise removal sounds worse. Accept some residual noise rather than introducing musical noise.

Practical Tips for Better Noise Reduction Results

Record a noise floor sample

Before your recording session, record 2–5 seconds of silence (just the room ambience with no one talking). This gives the noise reducer a perfect noise profile to work from.

Apply noise reduction first

In your processing chain, apply noise reduction before EQ, compression, or limiting. Noise reduction on an already-compressed signal can create more artifacts.

Use the minimum needed amount

The goal is to reduce noise to an acceptable level, not eliminate it completely. Over-aggressive noise reduction sounds worse than light noise. Use the lowest setting that makes the noise inaudible.

Monitor with headphones

Speakers can hide subtle artifacts. Use headphones to check for musical noise, especially in the silences between sentences.

Combine with high-pass filtering

A high-pass filter at 80–120 Hz removes low-frequency rumble before noise reduction runs. This prevents the algorithm from working on sub-bass noise it cannot easily model.

Use lossless source files

MP3 compression artifacts (metallic ringing, pre-echo) can confuse noise reduction algorithms. Use WAV or FLAC source files when possible for best results.

When to Use Noise Reduction

Podcasts & voiceovers

Remove HVAC hum and room noise from speech recordings. Even a modest 10–15 dB noise reduction dramatically improves perceived quality.

Video production

Fix on-location recordings with wind noise, traffic, or room tone. Clean dialogue is essential for professional-sounding video.

Music recordings

Remove amp hum, ground noise, and cable hiss from instrument recordings. Particularly useful for acoustic guitar, vocals, and vintage equipment.

Audio restoration

Remove tape hiss and vinyl crackle from archival recordings. Noise reduction is a standard part of the vinyl-to-digital transfer workflow.

Conference & call recordings

Clean up noisy remote participants in meeting recordings. Remove the constant background hum of home offices and open-plan environments.

Mastering

Reduce the noise floor of a mix before mastering, especially for quiet passages that will be amplified. Noise becomes more audible after limiting.

Frequently Asked Questions

How does noise reduction work in audio?

Audio noise reduction analyzes a silent section of the recording to build a noise profile, then subtracts that profile from every frame of the audio using spectral subtraction. The result is the signal with the consistent background noise removed, leaving a cleaner recording.

What is spectral subtraction noise reduction?

Spectral subtraction (Boll 1979) is the foundational algorithm for noise reduction. It estimates the noise spectrum from a silent section, then for each frequency bin subtracts the noise magnitude: |S(k)| = max(|X(k)| - α|N(k)|, β|X(k)|). The original signal phase is preserved to avoid artifacts.

Can I remove background noise from audio online for free?

Yes. WavinTools offers a free online noise reducer powered by spectral subtraction. It runs entirely in your browser — no upload, no account required. Upload any MP3 or WAV file and download the cleaned audio instantly.

What types of noise can be removed?

Noise reduction works best on stationary noise: room hum, HVAC noise, fan noise, electrical hiss, and tape hiss. It works less well on non-stationary noise like traffic or voices in the background, because these sounds change over time and cannot be modeled by a fixed noise profile.

What are the side effects of noise reduction?

Over-aggressive noise reduction can cause "musical noise" — a metallic, warbling artifact. This is reduced by using smooth Wiener masks, keeping the spectral floor above zero, and using the minimum oversubtraction needed to make noise inaudible.

Does noise reduction affect audio quality?

Light noise reduction (10–15 dB) on well-recorded audio has minimal impact on quality. Aggressive settings can introduce musical noise or slightly dull the high frequencies. The key is using the minimum amount needed to reduce noise to an acceptable level.

Try the Noise Reducer

Free Online Noise Reducer
Remove background noise using spectral subtraction — no upload, 100% free

Related Articles