Electrical Hum Removal — User Guide

What this does

Electrical Hum Removal is designed for stationary power-line hum: a fundamental near 50 Hz or 60 Hz together with integer harmonics such as 100/120 Hz, 150/180 Hz, and so on. It can detect the mains family automatically or work at a fixed 50 Hz or 60 Hz base frequency.

The processor does not apply a broad notch-filter bank. Instead, it estimates the amplitude and phase of the coherent sinusoidal component at each selected harmonic and subtracts that fitted sinusoid from every channel. Frequencies between the harmonic lines are left untouched by the cancellation stage.

This is a restoration tool for stable tonal hum, not a general broadband denoiser. It is most effective when the unwanted hum keeps approximately constant frequency, amplitude, and phase over the analysed portion of the recording.

Quick start

1. Select exactly one Sound object in Praat.
2. Run Electrical_Hum_Removal.praat.
3. For an unknown mains frequency, start with Auto-detect (mild).
4. If you know the recording contains 50 Hz or 60 Hz hum, choose the corresponding fixed preset or detection mode.
5. Use Dry_wet_mix as the cancellation depth: 0 leaves the sound unchanged; 1 subtracts the full fitted component.
6. Leave Output_level_mode at natural level unless you want a safety ceiling or explicit peak normalization.
7. Click OK. The processed Sound remains selected and can optionally be played immediately.

Presets

Preset values override the corresponding detection mode, harmonic count, and cancellation-depth settings.

Parameters

How it works

1. Hum-frequency selection

Fixed modes use exactly 50 Hz or 60 Hz. Auto mode analyses up to the first 10 seconds on a real input channel rather than a mono sum; for multichannel material it uses the channel with the highest RMS.

The detector examines a high-resolution, Hann-windowed spectrum and scores a harmonic comb around the 50 Hz and 60 Hz mains regions. A candidate must contain the fundamental and enough active harmonics above both the local spectral baseline and the global floor. The candidate is then refined from the measured harmonic peaks.

Auto mode also checks that the candidate lies close to a standard mains frequency and, on sufficiently long material, that the frequency remains stable across four sub-blocks. For very short files, a convincing candidate is snapped to the nearer 50 Hz or 60 Hz standard rather than trusting an imprecise measured value.

2. Coherent harmonic fitting

For each requested harmonic below 85% of Nyquist, the script fits a cosine/sine pair separately for every audio channel. The fit uses at most 10 seconds and is shortened to an integer number of mains cycles so that the harmonic frequencies align with the spectral analysis grid.

The fitted amplitude and phase are converted into a sinusoidal model and subtracted directly across the Sound. Each channel is estimated and processed independently, while the same base frequency and harmonic series are used across channels.

3. Cancellation depth

Dry_wet_mix is not a conventional dry/wet blend between two complete signals. It scales the amount of the fitted hum model that is subtracted. At 100%, the full estimated coherent component is removed; at lower values, part of that stationary component is deliberately left in place.

4. Output level

None (natural level) applies no global gain. Safety ceiling attenuates only when the resulting peak exceeds Peak_level. Peak normalize always scales a non-silent result so its peak reaches Peak_level.

Visualization

The Picture window provides a diagnostic view of what the processor measured and changed:

• Original spectrum: LTAS up to 1 kHz (or Nyquist if lower), with vertical markers at the fitted cancellation frequencies.
• Processed spectrum: measured before the output-level stage and shown on the same dB scale as the original.
• Waveforms: the first 0.5 seconds of original and processed audio on one shared amplitude scale.
• Cancellation table: harmonic number, fitted frequency, estimated tone level, and cancellation percentage; up to eight entries are shown explicitly.
• Summary strip: base frequency, number of harmonics, fit duration, cancellation depth, peak before/after the level stage, and processing time.

The visualization is diagnostic only; it does not change the audio.

Output behavior

• Input requirement: exactly one Sound object.
• Channels: preserved; every channel is processed independently.
• Sample rate: preserved.
• Duration: preserved.
• Start time: preserved. Processing uses a temporary time-zero copy and restores the original xmin before returning the result.
• Randomness: none. The same input and settings are deterministic.
• Bypass: Dry_wet_mix = 0 returns an exact copy without cancellation or output-level processing. In Auto mode, “Return the input unchanged” also bypasses the cancellation when no convincing hum is found.
• Output name: <original>_hum_removed_<preset>.
• Peak warning: with natural level, a peak above 1.0 is reported because it can clip when saved to integer PCM.