Repairing Broadband Noise and Hum

Broadband noise and mains hum are different defects. A denoiser estimates a time-varying spectrum and applies a gain mask. A dehummer tracks a fundamental and its harmonics. Use the detector for the defect you actually hear, and check that a sustained musical tone is not being mistaken for noise.

Terms on this page

• STFT (short-time Fourier transform): the signal cut into overlapping frames of nFft samples, hopLength apart, each turned into a spectrum of frequency bins. See Spectrogram and STFT.
• Gain mask: one gain between a floor and 1 per bin and frame; the denoiser multiplies the spectrum by it and converts back to audio.
• Noise estimator: how the noise level in each bin is worked out. quantile averages the quietest 10% of frames into one fixed spectrum; mcra, imcra and spp track the noise frame by frame and can follow a changing floor. MCRA and IMCRA (minima-controlled recursive averaging) follow the floor through the spectral minima; SPP weights each update by the probability that speech is present.
• MMSE-STSA / LogMMSE: gain rules that set each bin's gain from its estimated signal-to-noise ratio; spectral subtraction subtracts the noise power instead. The decision-directed smoothing (ddAlpha, which blends each frame's gain with the previous frame's estimate) keeps the gains from flickering into "musical noise": isolated bins that randomly survive the mask and twinkle like faint tones.
• Fundamental and harmonics: mains hum is a tone at 50 or 60 Hz plus whole-number multiples of it. subtract mode cancels each one with a tracked sine; notch mode filters each one out with a narrow notch. Q is the centre frequency divided by the bandwidth, so Q 20 at 50 Hz is a 2.5 Hz-wide notch; PLL (phase-locked loop) is the tracker that follows a drifting mains frequency.

Choose the stage

Defect	Detect with	Repair with	Main risk
Steady hiss, mic floor, or tape noise	masteringRepairDetectNoiseFloor	masteringRepairDenoiseClassical	Quiet musical material can be pulled down with the estimated floor.
A 50/60 Hz tone and harmonics	masteringRepairDetectHum	masteringRepairDehum	A notch can remove intentional content at the same frequencies.

The hum detector only reports a stationary line: a repeating kick reads fundamentalProminence 1.0, a linear ratio where 1.0 means no peak (it is not a lock flag). Neither stage is a loudness processor. Neither separates a source from a finished mix, and neither repairs clicks, clipping, or room reverb.

Detect first, then apply

The request-object form makes the sample rate and options explicit. decodedMonoPcm must be a Float32Array. A buffer shorter than nFft is padded for the analysis and comes back at its own length.

import {
  init,
  masteringRepairDenoiseClassical,
  masteringRepairDehum,
  masteringRepairDetectHum,
  masteringRepairDetectNoiseFloor,
} from '@libraz/libsonare';

await init();

const sampleRate = 48_000;
const samples = decodedMonoPcm;
const noise = masteringRepairDetectNoiseFloor({
  samples,
  sampleRate,
  nFft: 1024,
});
const hum = masteringRepairDetectHum({ samples, sampleRate });

console.log(noise.floorDbfs, noise.bandFloorDbfs);
console.log(hum.fundamentalHz, hum.harmonics, hum.harmonicDbfs);

// Keep only the stages supported by the reports and an audition.
const dehummed = masteringRepairDehum({
  samples,
  sampleRate,
  adaptive: true,
});
const cleaned = masteringRepairDenoiseClassical({
  samples: dehummed,
  sampleRate,
  reductionDb: 26,
});

The example follows the fixed repair order: dehum precedes denoise. In an application, make each call conditional on the report and the listening decision rather than treating a positive measurement as permission to remove all of it.

Denoise settings

masteringRepairDenoiseClassical is an STFT-domain classical denoiser. Its options are the public TypeScript DenoiseClassicalOptions; use the API reference for the complete overload list.

Option	Values or default	Use
mode	logMmse (default), mmseStsa, spectralSubtraction	Select the gain function. Spectral subtraction also reads overSubtraction and spectralFloor.
noiseEstimator	quantile (default), mcra, imcra, spp	Select how the noise spectrum is tracked. A chain override takes the same names (repair.denoise.noiseEstimator, repair.denoise.mode, repair.dehum.mode). The streaming mastering chain defaults to spp and refuses quantile by name; the capability catalog describes that realtime insert, so it lists spp as the default and leaves out quantile.
nFft / hopLength	1024 / 256 by default	Set the STFT window length and hop. See the constraints below.
reductionDb	26 by default, finite and in [0, 120]	Set the deepest attenuation of any gain-mask bin. It is a residual-noise floor, not a target loudness.
noiseEstimationQuantile	0.1 by default	Fraction of the quietest frames used by the quantile estimator.
ddAlpha, speechPresenceGain, gainSmoothing	See the API type	Control decision-directed smoothing and the MMSE gain path.

At reductionDb: 26, the mask floor is a linear gain of about 0.0501, or a maximum per-bin attenuation of 26 dB; it does not promise that the broadband noise itself becomes 26 dB quieter. A larger value allows deeper attenuation, but it does not make the estimate more reliable. The report's maxReductionDb and floorLimitedFraction show whether this bound controlled the result. The figure below shows the bound on one frame.

One frame through the gain mask

• noisy frame
• estimated noise
• gain per bin

A harmonic tone over a sloped noise floor, one 1024-point frame at 48 kHz. Top: the frame and the estimated noise. Bottom: the gain the default logMmse rule gives each bin. Where a harmonic stands well clear of the noise the gain stays near 0 dB. Between the harmonics and above them, where the bin is only noise, the gain falls until it reaches −reductionDb (−26 dB here) and stops there. Bins whose noise happens to read high in this frame keep a little more.

nFft and hopLength follow the STFT rules shared by every repair stage, which StreamingDenoise (including the StreamingMasteringChain live path) also applies. Streaming additionally requires a recursive estimator (mcra, imcra, or spp).

Dehum settings

masteringRepairDehum removes the fundamental and its harmonics. Every option has an upper bound; a larger value throws InvalidParameter.

Option	Default	Accepted	Use
fundamentalHz	50	(0, 5000]	Mains frequency to remove; set 60 where the supply is 60 Hz.
harmonics	4	[1, 16]	Number of harmonics, fundamental included.
mode	subtract	subtract, notch	Cancel the tracked sinusoids, or filter with notches.
q	20	(0, 100]	Selectivity of each canceller or notch (centre frequency ÷ bandwidth; larger is narrower).
adaptive	false	boolean	Search for the actual fundamental instead of trusting fundamentalHz.
searchRangeHz	2	[0, 100]	Half-width of the adaptive search around fundamentalHz.
adaptation, pllBandwidth	0.25, 0.01	[0, 1]	Adaptive tracking speed; pllBandwidth is the bandwidth of the phase-locked loop that follows the fundamental.
frameSize	2048	[16, 16384]	Analysis frame of the adaptive tracker.

The hum detector and the adaptive search read the window fundamentalHz ± searchRangeHz; a window that lies entirely outside [1 Hz, 0.49 × sample rate] is refused (it is not clamped into range). The same holds for the linked and streaming forms.

In subtract mode each harmonic canceller starts from a joint least-squares fit of all the harmonic references over the opening window instead of from zero, so a short clip is attenuated from its first samples. When the window is too short to tell the harmonics apart, the cancellers are left unseeded. On a 0.25 s clip carrying 50 Hz and 100 Hz hum under a 440 Hz tone at 22.05 kHz, the RMS of the first 1000 samples fell from 0.1675 to 0.0355, which is the tone alone.

A/B PROCESS · DEHUMIDLE

Dehum repair — mains hum before and after

Scroll to load

A 50 Hz hum with three harmonics is added to a clean chord, and the spectrum is zoomed to 0 to 1 kHz so the four lines are visible. Compare the damaged and repaired signals, then change how the hum is removed. HUM scores the mean level between 40 and 250 Hz: cancelling or notching all four components lowers it by about 11 dB, while cancelling only the fundamental lowers it by about 2 dB and leaves the 100, 150 and 200 Hz harmonics. No loudness matching is applied.

Compare

Removal

Channel handling

Use the mono entry point for one channel, the stereo entry point for a left/right pair, and the linked entry point for any number of channels with one shared mask.

import {
  masteringRepairDenoiseClassicalLinked,
  masteringRepairDenoiseClassicalStereo,
} from '@libraz/libsonare';

const stereo = masteringRepairDenoiseClassicalStereo({
  left,
  right,
  sampleRate,
  reductionDb: 26,
});

const linked = masteringRepairDenoiseClassicalLinked({
  channels: [frontLeft, frontRight, center],
  sampleRate,
  reductionDb: 26,
});

The stereo and linked forms build the mask from summed channel power and apply it unchanged to every channel. This keeps interchannel level and phase differences from moving. Their report has one pair- or set-level measurement. floorDbfs is absolute and rises by about 3.01 dB for two identical channels and 4.77 dB for three, so compare it only with a measurement made over the same channel count. All input channels must have the same length; one shorter than nFft is padded for the analysis.

masteringRepairDehumStereo keeps independent filter or canceller state per channel, but with adaptive: true it tracks one fundamental for both. It chooses the channel with the stronger hum as the tracking reference, searches from the summed projected energy of both original channels, and sends one tracked fundamental to both paths. Hum on only one side and opposite-phase hum therefore remain visible to the tracker, while each channel's amplitude and phase are handled by its own state. With the default adaptive: false, each channel filters the configured frequency independently.

Read the result

masteringRepairDetectNoiseFloor reports floorDbfs and a 32-band bandFloorDbfs shape before any gain mask is applied. It does not report attenuation. masteringRepairNoiseBandBins({ nFft, sampleRate }) returns the 33 bin edges of those 32 bands, so band k spans bins bins[k] to bins[k + 1]; a band whose edges coincide is empty, and its level is a fixed placeholder value, not a measurement.

The mono denoise call returns only the samples. For the attenuation numbers, call masteringRepairDenoiseClassicalStereo or masteringRepairDenoiseClassicalLinked (a one-channel linked call works) or masteringRepairApply. Their report contains:

• meanReductionDb is the mean attenuation of the mask;
• maxReductionDb is the deepest attenuation applied;
• floorLimitedFraction is the fraction of mask cells that reached reductionDb.

These values describe what the processor did. They are not LUFS matching and they do not prove that the remaining noise is inaudible. Measure integrated loudness separately if the delivery target requires it.

Listen to a denoise A/B

A/B PROCESS · DENOISEIDLE

Denoise repair — damaged vs repaired

Scroll to load

Broadband hiss is added to a clean chord. Compare the noisy and denoised signals and their spectra, then choose a denoising algorithm. Both retain their processed levels without extra loudness matching. Noise reduction can also change musical transients and timbre. FLOOR shows the average high-band reduction in dB.

Compare

Algorithm

This A/B uses the classical denoiser and applies no loudness matching, so changes in noise and timbre remain audible.

Limits

The denoiser estimates noise from the buffer itself. A calibration tone, drone, or long held note can be treated as noise by the estimator and attenuated along with the floor. Classical denoise is not a learned model either: speech-enhancement networks generally remove noise from voice more cleanly, especially noise that changes over time, while this denoiser is deterministic and needs no model weights. See Signal processing versus learned models. The detector and classical helper calls are offline; StreamingDenoise and StreamingMasteringChain provide block processing for denoise with mcra, imcra, or spp, while the whole-signal quantile estimator is rejected. See Audio Repair Workflow for the shared order and JavaScript mastering API for exact report types.