Spectral Subtraction
Definition
An enhancement method that estimates the noise spectrum during a speech-free interval and subtracts it from every subsequent frame. Simple and effective for stationary noise, but prone to musical noise artefacts when the noise estimate is imperfect.
- Field
- Audio forensics, speech enhancement
- Method
- Estimate noise in speech-free interval, subtract from later frames
- Effective for
- Stationary noise
- Known artefact
- Musical noise from imperfect noise estimate
Common questions
Why does spectral subtraction struggle with non-stationary background noise?+
The method assumes the noise profile measured in a quiet interval still applies later in the recording, so noise that changes character over time, such as traffic or wind, is only partially removed and can leave residual artefacts.
How does musical noise differ from the original background noise?+
Musical noise appears as scattered tonal blips introduced by the subtraction process itself, not as a feature of the original recording, and can mislead a listener judging authenticity if mistaken for genuine signal content.
Related terms
- Non-Stationary Noise
- Interference whose spectral content changes over time: a vehicle passing, a crowd noise that rises and falls, or a door slamming. Adaptive...
- PESQ / STOI
- Perceptual Evaluation of Speech Quality and Short-Time Objective Intelligibility: instrumental measures that predict listener intelligibility without needing human subjects. STOI scores between...
- Stationary Noise
- Noise whose spectral characteristics remain approximately constant over the duration of a recording. Tape hiss, HVAC fan noise, and mains hum are...
- SWGDE
- Scientific Working Group for Digital Evidence. A US multi-agency body that publishes consensus best-practice documents for digital forensic disciplines, including image authentication,...
- Wiener Filter
- A minimum mean-squared-error filter that computes a frequency-dependent gain function from the signal-to-noise ratio in each frequency band. Less aggressive than spectral...