Noice reduction strength

[comicdodge]

Is there a way to change the reduction strength? Although for a lot of cases it works properly in my domain (singing), some parts are very bad and disturb my signal. I'd like to use the model but to also control how much we should integrate

[shanto12]

This is a common request for deep learning-based noise reduction models. The challenge is balancing aggressive noise reduction with preserving the desired signal (vocals in your case).

Implementation Options:

1. Attenuation/Mix Parameter: Add a configurable parameter (0-1) that blends the processed output with the original signal. This gives users control without retraining the model.

2. Threshold-Based Processing: Implement a noise gate threshold so only segments above a certain SNR level get processed, leaving cleaner sections untouched.

3. Multi-Model Approach: Provide different model variants (light/medium/aggressive) trained with different noise reduction targets.

4. Post-Processing EQ: Allow users to apply frequency-selective attenuation to compensate for over-processing in specific bands.

For singing specifically, consider that DeepFilterNet might be over-suppressing breath sounds and vocal texture that are musically important. A "vocal preservation" mode with lighter processing could help maintain the natural character while reducing background noise.