Recording Noise Removal: Two Engines, Three Strengths
Record a lecture at home or a meeting in a conference room and the playback reveals the truth: steady noise everywhere. A denoiser can rescue the track, but the wrong engine or too much strength damages the voice itself. Two engines, three strengths — here is how to combine them.
First, classify the noise
Denoisers excel at steady noise with a stable spectrum: HVAC hum, electrical hiss, fans, constant buzz. Transient noises — keyboard clatter, door slams, coughs — are out of scope; trim those instead. Only run denoising when steady noise is actually present.
Choosing between the two engines
The AI engine (RNNoise) separates voice from steady noise better — first choice for speech. The spectral engine is the classic method, better for non-voice material like instruments and ambience, or as a comparison when AI underperforms. Rendering both and listening is the fastest way to the best result.
- Speech content: AI engine first
- Non-voice material: spectral engine first
- Unsure: render both and compare
Strength: lighter is safer
Light keeps the most detail and suffices for mild noise; standard is the recommended starting point; strong is for badly polluted material. Over-processing has a specific sound: muffled voice, ghosting tails, watery artifacts — and it is irreversible. Start at standard and step down if it already sounds clean.
After denoising
Clean voice is ready for loudness normalization and export in one pass. Denoise once — repeating on the same material compounds damage and only gets worse.