Python Noise Reduction — Core Concepts

What noise reduction does

Audio noise reduction removes unwanted background sounds — hiss, hum, fan noise, environmental ambience — from a recording while preserving the desired signal (speech, music, instruments). It is essential for podcasting, transcription preprocessing, music production, and any application where audio quality matters.

Types of noise

Type	Example	Character
Stationary	Fan hum, tape hiss, electrical buzz	Constant, predictable spectrum
Non-stationary	Traffic, crowd chatter, keyboard clicks	Changes over time
Impulse	Clicks, pops, crackle	Short, sharp bursts

Stationary noise is the easiest to remove. Non-stationary noise requires adaptive methods. Impulse noise needs specialized detection and interpolation.

Spectral gating — the core technique

Spectral gating is the most common classical approach:

1. Estimate the noise spectrum. Analyze a noise-only segment (or estimate adaptively) to get the average magnitude per frequency bin.
2. Compute a threshold. For each frequency bin, set a gate threshold based on the noise estimate.
3. Apply the gate. For each time-frequency cell in the STFT, if the magnitude is below the threshold, attenuate it; if above, pass it through.
4. Reconstruct. Inverse STFT converts the cleaned spectrogram back to audio.

The gate is usually “soft” — it reduces noise smoothly rather than cutting it to zero, which would create artifacts (musical noise).

The noisereduce library

The simplest way to do noise reduction in Python:

import noisereduce as nr
import soundfile as sf

audio, sr = sf.read("noisy_recording.wav")

# With a noise sample (best quality)
noise_sample = audio[0:sr]  # first second is noise-only
cleaned = nr.reduce_noise(y=audio, sr=sr, y_noise=noise_sample)

# Without a noise sample (automatic estimation)
cleaned = nr.reduce_noise(y=audio, sr=sr)

Key parameters

• prop_decrease (0.0–1.0): How aggressively to remove noise. 1.0 removes everything below the threshold; 0.5 reduces by half. Default 1.0.
• stationary: If True, assumes noise is constant. If False, uses adaptive estimation (better for changing noise).
• n_fft: FFT window size. Larger values give better frequency resolution but worse time resolution.
• freq_mask_smooth_hz and time_mask_smooth_hz: Smooth the noise gate across frequency and time to reduce artifacts.

Practical workflow

1. Load audio with soundfile or librosa
2. Identify noise segment — a silent pause, intro/outro without speech
3. Apply noise reduction with conservative settings first
4. Listen to the result — check for artifacts (metallic sound, underwater effect)
5. Adjust — if too aggressive, lower prop_decrease; if noise remains, increase it
6. Export with soundfile.write()

Artifacts and tradeoffs

Aggressive noise reduction causes:

• Musical noise — random tonal artifacts from the gate opening and closing at individual frequency bins
• Loss of transients — quiet consonants (s, t, f) can be mistaken for noise
• Hollow sound — over-suppression removes harmonics along with noise

The rule of thumb: it is better to leave a little noise than to destroy the signal. Start conservative and increase reduction incrementally.

Common misconception

Noise reduction cannot recover information that noise has destroyed. If a frequency band is completely masked by noise, removing the noise reveals silence — not the hidden signal. Noise reduction works best when the signal-to-noise ratio is already moderate (the desired sound is louder than the noise).

How it fits with other tools

For preprocessing: apply noise reduction before feeding audio to Whisper (speech recognition) or Librosa (feature extraction) — cleaner input improves downstream accuracy. For real-time denoising, combine spectral gating with sounddevice streaming. For deep-learning denoising, use models like DTLN or RNNoise (see Deep Dive).

One thing to remember: Spectral gating estimates the noise spectrum and attenuates matching frequencies in the recording — the noisereduce library wraps this entire pipeline into a single function call with sensible defaults.