Removing Noise from Video Audio: Hiss, Wind and Hum
Noise reduction trades sound quality for cleanliness, and the exchange rate depends on noise type. Steady noise cleans up beautifully; transient noises fight back. Diagnose first, then pick engine and strength.
Noise types and what to expect
- Steady hiss / device noise: uniform and constant — easiest, dramatic improvement.
- Wind rumble: low-frequency whoosh that moves with speech — needs medium strength.
- Electrical hum: fixed-frequency buzz from AC/fans — easy, the frequency never moves.
- Transient noise (doors, keyboards): overlaps speech spectrum — reduction hurts the voice.
Two engines
AI (RNNoise-class neural): best for non-stationary noise like wind and varied environments, high voice fidelity, slower.
Classic FFT spectral: fast and surgical on steady hiss and fixed hum; leaves watery artifacts on complex noise.
Strength discipline
Light → Standard → Strong: start low and step up only as needed. Over-reduction produces watery, robotic voices and lost breaths — worse than the mild hiss you started with.
After reduction, loudness often reads quiet (the noise floor was part of the level) — normalize afterward. Reduced noise also lifts transcription accuracy.