About

Background noise degrades audio intelligibility and listener retention. A recording with a signal-to-noise ratio (SNR) below 20 dB loses roughly 30% of perceived clarity. This tool applies spectral subtraction: it samples a noise-only segment to build a frequency-domain noise profile, then subtracts that profile's magnitude spectrum from each frame of the full signal. A noise gate zeros out samples below a configurable amplitude threshold T with attack/release smoothing to prevent audible clicks. Multi-band equalization attenuates persistent tonal noise that spectral subtraction may miss. All processing runs locally in your browser via OfflineAudioContext. No audio data leaves your machine.

This tool approximates studio noise reduction under the assumption that the noise is roughly stationary (constant hum, fan noise, hiss). Non-stationary noise such as speech crosstalk or impact sounds will not be cleanly separated. For best results, select a segment of pure noise lasting at least 0.5 s. Over-aggressive subtraction introduces "musical noise" artifacts. Start with moderate settings and increase gradually.

Formulas

The core technique is spectral subtraction. For each short-time FFT frame of length N, the estimated clean signal spectrum is computed as:

Ŝ(k) = |X(k)| − α ⋅ |N̂(k)|

where X(k) is the noisy signal's DFT at bin k, N̂(k) is the averaged noise profile magnitude, and α is the subtraction strength (over-subtraction factor, typically 1.0 - 4.0). A spectral floor β prevents negative magnitudes:

|Y(k)| = max(Ŝ(k), β ⋅ |X(k)|)

where β ∈ [0.001, 0.1] controls the residual noise floor. The phase of the original signal is preserved. The noise gate applies an amplitude envelope:

g(n) =

{

1 if |x(n)| ≥ T0 if |x(n)| < T

where T is the threshold amplitude and g(n) is smoothed with exponential attack/release to avoid clicks. Signal-to-Noise Ratio improvement is measured as:

SNR = 20 ⋅ log₁₀ RMS_signalRMS_noise dB

where RMS_signal is the root-mean-square of the desired signal and RMS_noise is the RMS of the noise floor. α = subtraction strength. β = spectral floor coefficient. T = noise gate threshold. k = frequency bin index. N = FFT size.

Reference Data

Noise Type	Typical Frequency Range	Recommended Strategy	Spectral Subtraction Strength	Noise Gate Threshold
AC Hum (50 Hz)	50 - 150 Hz	Highpass filter + notch at 50 Hz	1.0 - 2.0	−50 dB
AC Hum (60 Hz)	60 - 180 Hz	Highpass filter + notch at 60 Hz	1.0 - 2.0	−50 dB
Fan / HVAC Noise	100 - 500 Hz	Spectral subtraction	1.5 - 3.0	−45 dB
White Noise / Hiss	1 - 20 kHz (broadband)	Spectral subtraction + gentle lowpass	2.0 - 4.0	−40 dB
Pink Noise	20 - 20000 Hz (1/f)	Spectral subtraction	2.0 - 3.5	−42 dB
Electrical Buzz	100 - 1000 Hz	Notch filter + spectral subtraction	1.5 - 2.5	−48 dB
Computer Fan	200 - 2000 Hz	Spectral subtraction + EQ cut	2.0 - 3.0	−44 dB
Traffic Rumble	20 - 300 Hz	Highpass at 80 - 120 Hz	1.0 - 2.0	−50 dB
Wind Noise	20 - 500 Hz	Highpass at 100 Hz + compression	1.5 - 2.5	−46 dB
Camera Hiss	4000 - 16000 Hz	Lowpass filter + mild spectral subtraction	1.0 - 2.0	−38 dB
Fluorescent Light Buzz	120 - 240 Hz	Notch filter at harmonics	1.5 - 2.5	−48 dB
Microphone Self-Noise	1000 - 8000 Hz	Spectral subtraction + noise gate	2.0 - 3.5	−40 dB
Ground Loop Hum	50/60 Hz + harmonics	Multi-notch filter	1.0 - 2.0	−52 dB
Room Ambience / Reverb Tail	200 - 5000 Hz	Noise gate + mild spectral subtraction	1.0 - 1.5	−36 dB
Tape Hiss (Analog)	2000 - 16000 Hz	Spectral subtraction + de-emphasis EQ	2.5 - 4.0	−38 dB

Frequently Asked Questions

A minimum of 0.5 seconds of pure noise is recommended. Longer samples (1-2 seconds) produce a more stable average noise spectrum, reducing variance in the estimate. The noise segment must contain only the noise you want to remove - no speech, music, or transient sounds. If the segment is too short (under 0.3 s), the FFT averaging will have insufficient frames, and the resulting profile will be noisy itself, leading to artifacts.

Musical noise (also called "birdies" or 'twinkling') occurs when spectral subtraction removes noise inconsistently across frequency bins from frame to frame, leaving isolated spectral peaks that sound like random tones. This is caused by over-aggressive subtraction strength (α > 3.0) combined with a low spectral floor (β < 0.01). To minimize it: reduce the subtraction strength to 1.5-2.5, increase the spectral floor to 0.02-0.05, and apply more smoothing frames. A small amount of residual noise is preferable to musical artifacts.

No. Spectral subtraction assumes noise is statistically stationary - meaning its spectral characteristics remain constant over time. Background speech, door slams, or variable-speed fans violate this assumption. The noise profile captured from one segment will not match the noise in other parts of the recording. For non-stationary noise, adaptive filtering or neural-network-based separation would be required, which is beyond the scope of this browser-based tool.

FFT size (N) determines frequency resolution: Δf = sample_rate / N. With a sample rate of 44100 Hz and N = 2048, each bin covers approximately 21.5 Hz. Larger FFT sizes (4096, 8192) provide finer frequency resolution, allowing more precise noise profile matching, but increase processing time and reduce temporal resolution (the ability to track rapid changes). For most speech recordings, N = 2048 or 4096 provides a good balance. For music with sustained tones, N = 4096 or 8192 may be preferable.

They operate on different domains. Spectral subtraction works in the frequency domain (per-bin magnitude reduction), while the noise gate works in the time domain (sample-level amplitude gating). Apply spectral subtraction first to remove steady-state noise, then use the noise gate to silence residual low-level noise during pauses. Setting the gate threshold too high will clip quiet speech. A threshold of −40 to −50 dB relative to peak is typical for voice recordings. Always use attack/release smoothing (5-20 ms) to avoid audible clicks at gate transitions.

Yes, to some degree. All noise reduction is a trade-off between noise attenuation and signal distortion. Aggressive subtraction removes noise but also removes signal components that overlap spectrally with the noise. This manifests as a slightly "hollow" or "underwater" quality in speech. The spectral floor parameter β controls how much residual noise is retained to mask this distortion. For critical applications, always compare the processed output against the original using the A/B toggle and use the minimum subtraction strength that achieves acceptable noise reduction.

This tool accepts any format decodable by the Web Audio API in your browser, which typically includes WAV, MP3, OGG, FLAC, AAC, and WebM audio. The decoded audio is processed at its native sample rate (commonly 44100 or 48000 Hz). Output is exported as WAV (PCM 16-bit) to preserve quality. The maximum file size supported is 50 MB. Files exceeding this may cause memory issues in the browser tab.