Harmonic Remover — v1.0 User Guide

Selectively removes the fundamental frequency and/or chosen overtones from a sound. Uses pitch tracking and a bandpass subtraction method (Hann‑windowed frames, 75 % overlap) to isolate and subtract harmonic content, leaving the residual (noise, breath, or remaining partials). Ideal for spectral sculpting, debuzzing, noise extraction, and educational demonstrations.

Author: Shai Cohen Affiliation: Department of Music, Bar‑Ilan University, Israel Version: 1.0 (2025) License: MIT License Repo: GitHub

Contents:

What it does Quick start Presets (6 archetypes) Method: bandpass subtraction Parameters Visualization FAQ / troubleshooting

What this does

Harmonic Remover extracts the pitched (harmonic) content from a sound and subtracts it, leaving the residual (noise, breath, or remaining partials). The algorithm:

Detects the fundamental frequency (F0) per frame using Praat’s pitch tracker.
For each voiced frame, isolates each harmonic (fundamental and/or selected overtones) using a Hann bandpass filter that captures the full spectral spread (including vibrato).
Subtracts the isolated harmonic bands from the original frame.
Reconstructs the output via overlap‑add (75 % overlap, exact COLA normalisation).

Why bandpass subtraction instead of a spectral notch? A narrow spectral notch removes only the exact centre frequency, missing vibrato or pitch variation within a frame. Bandpass isolation captures the entire harmonic region (width = ±notch_bandwidth_Hz), then subtracts it, leaving the surrounding noise and other partials intact.

Quick start

In Praat, select exactly one Sound object (mono or stereo – stereo will be downmixed to mono for processing).
Run script… → Harmonic_Remover.praat.
Choose a Preset:
- Remove Fundamental, Remove Odd Harmonics, Remove Even Harmonics, Remove All Overtones (keep F0 only), Remove F0 + All Overtones (noise/breath only), Remove High Harmonics (warm/mellow)
For custom mode (preset = Custom), adjust parameters as desired:
- Removal_mode – Fundamental only / Selected overtone numbers / All overtones (keep fundamental) / Fundamental + all overtones.
- Overtone_numbers – comma‑separated list (e.g., 2,4,6) for selected mode.
- Pitch_floor / Pitch_ceiling – range for F0 detection.
- Number_of_harmonics – up to 20 partials considered.
- Analysis_window_ms – frame length (20–200 ms).
- Notch_bandwidth_Hz – half‑width of the bandpass filter around each harmonic (± value).
- Min_amplitude_dB – skip frames below this RMS (avoid amplifying noise).
- Wet_dry_percent – blend between processed (wet) and original (dry).
Click OK. The script analyses pitch, processes frames, and creates a new Sound object named originalname_harmonicRemoved_preset.

Tip: Start with Remove Fundamental to hear the “de‑buzzed” effect (useful for removing hum). Remove All Overtones (keep F0 only) produces a sine‑like tone. Remove F0 + All Overtones extracts the noise/breath residual.

Important: This effect is implemented entirely in Praat – no Python required. Processing time depends on the number of frames and harmonics removed; for long files (e.g., 10 s, 2048 frames, 12 harmonics each), it may take 10–30 seconds.

The 6 presets (+ Custom)

PresetRemoval modeTarget harmonicsResult Remove FundamentalFundamental only1st harmonicRemoves the fundamental, leaving overtones – “hollow” or “bell‑like” texture. Remove Odd HarmonicsSelected overtones3,5,7,9,11Removes odd harmonics (square‑wave‑like filtering). Remove Even HarmonicsSelected overtones2,4,6,8,10Removes even harmonics – results in an “octave‑up” or hollow quality. Remove All Overtones (keep F0 only)All overtones (keep fundamental)2…NOutputs a near‑sinusoidal tone at the fundamental frequency. Remove F0 + All Overtones (noise/breath only)Fundamental + all overtones1…NExtracts the residual – breath, noise, or unvoiced content. Remove High Harmonics (warm/mellow)Selected overtones4…12Keeps harmonics 1‑3, removes higher partials – warm, mellow tone.

Method: bandpass subtraction

Processing pipeline:

Pitch detection – To Pitch with user‑defined floor/ceiling.
Framing – Hann window, 75 % overlap (hop = win/4). COLA property: sum of windows = 2.0 at every point.
For each voiced frame:
- For each harmonic to be removed (fundamental and/or selected overtones):
- Isolate the harmonic using Filter (pass Hann band) with centre = hn × F0, width = 2 × notch_bandwidth_Hz.
- Subtract the isolated band from the frame.
Unvoiced or low‑amplitude frames pass through unmodified (preserves noise).
Overlap‑add – after OLA, divide by 2.0 for exact unity gain.
Wet/dry mix – blend with original.

Hann bandpass filter is implemented with Praat’s built‑in Filter (pass Hann band), which uses a smooth, zero‑phase filter (no ringing). The transition width is set to notch_bandwidth_Hz / 2 to ensure the filter captures the full harmonic spread while not interfering with neighbouring partials.

📐 COLA normalisation

For a Hann window with 75 % overlap (hop = win/4), the sum of overlapping windows is exactly 2.0 at every sample. After OLA, the output is divided by 2.0 to restore unity gain. This is more accurate than peak‑based normalisation and preserves the original amplitude of non‑processed frames.

Parameters & defaults

Removal mode

Parameter	Options	Description
Removal_mode	Fundamental only / Selected overtone numbers / All overtones (keep fundamental) / Fundamental + all overtones	Which harmonic components to subtract.
Overtone_numbers	comma‑separated integers	Used only in “Selected overtone numbers” mode. Example: `2,4,6` removes the 2nd, 4th, and 6th partials.

Pitch detection

Parameter	Range	Default	Description
Pitch_floor	≥ 50 Hz	75 Hz	Lowest expected fundamental frequency.
Pitch_ceiling	≤ 2000 Hz	600 Hz	Highest expected fundamental frequency.
Number_of_harmonics	1–20	8	How many partials to consider (for mode “All overtones” and “F0+all”).

Spectral removal

Parameter	Range	Default	Description
Analysis_window_ms	20–200 ms	60 ms	STFT frame length. Larger = better frequency resolution, poorer time resolution.
Notch_bandwidth_Hz	5–200 Hz	60 Hz	Half‑width of the bandpass filter around each harmonic. Larger = more aggressive removal (captures vibrato).

Threshold & output

Parameter	Range	Default	Description
Min_amplitude_dB	-∞ … 0	-50 dB	Frames below this RMS level are not processed (avoid amplifying noise).
Wet_dry_percent	0–100	100	100 = fully processed; 0 = fully original.
Scale_peak	0–1	0.95	Normalise output peak to this value.
Draw_visualization	yes/no	yes	Show 5‑panel visualisation.
Play_result	yes/no	yes	Auto‑play after processing.

Visualization (5‑panel Praat picture)

When Draw_visualization = 1, the script draws:

Input and output waveforms – side‑by‑side.
Original and processed spectrograms (0–5 kHz).
Notch shape diagram – a frequency‑domain view showing which harmonics are removed (red shaded bands with “×”) and which are kept (green markers with numbers). The diagram is drawn at the median fundamental frequency of the source.
Summary panel – mode, number of harmonics, notch bandwidth, window, median F0, voiced frames, wet/dry, processing time.

Tip: The notch diagram is invaluable for verifying that you’ve selected the correct overtones. Red bands show exactly which frequency regions will be subtracted.

FAQ / troubleshooting

No pitched content detected / all frames unvoiced

Adjust Pitch_floor and Pitch_ceiling to match your source. For speech, 75–600 Hz is typical; for bass instruments, lower the floor to 40 Hz; for high soprano voices, raise the ceiling to 800 Hz. The script prints the median F0 in the Info window – if it’s 0, no pitch was found.

Output has clicks / artefacts

The OLA normalisation (division by 2.0) ensures perfect reconstruction for unprocessed frames. If clicks appear, try reducing Notch_bandwidth_Hz (sharper filters) or increasing Analysis_window_ms (smoother spectral transitions). Also ensure Min_amplitude_dB is not set too high – low‑amplitude frames are passed through unmodified, which can cause discontinuities if they are adjacent to heavily processed frames. A small wet/dry mix (e.g., 80 %) can smooth transitions.

Removal not effective for vibrato / pitch‑varying sounds

The bandpass filter width is centred on the instantaneous F0 per frame. If vibrato is very wide, increase Notch_bandwidth_Hz to capture the full frequency sweep. For very fast pitch changes (e.g., glissandi), consider using a smaller analysis window (20–30 ms) to better track the trajectory.

Why not use a spectral notch (FFT bin zeroing)?

Zeroing FFT bins introduces severe ringing artefacts (pre‑echo/post‑echo) because the abrupt truncation in the frequency domain creates a sinc‑shaped impulse response. The bandpass subtraction method uses a smooth, zero‑phase filter, preserving the time‑domain envelope of the harmonic while removing its energy. It also naturally handles vibrato and pitch variation within the frame.

Processing time

For a 10‑second sound at 44.1 kHz, analysis_window_ms = 60 gives about 665 frames. If removing 12 harmonics per frame, that’s ~8000 bandpass filter calls. Each filter is implemented in Praat’s C‑level code, so it’s reasonably fast (10–30 seconds). The progress indicator updates every 50 frames.