Microphone Noise Suppression DSP: Spectral Subtraction vs Deep Learning Filters: Edge Cases, Failure Modes, and Root-Cause Debugging

I’ve been stress-testing two noise suppression pipelines for competitive play—spectral subtraction (the classic Wiener-filter approach) versus a lightweight deep learning model like RNNoise or a custom inference engine—and I’m hitting real tradeoffs I can’t pin down. With spectral subtraction, I keep getting that underwater, muffled artifact during long vowel sounds and sudden loud key presses, but the latency stays under 20ms. The DL filter kills the mechanical click almost entirely and preserves my voice better, but I’m seeing 40-60ms latency spikes on burst traffic that wreck comms mid-round. My question: how are you all handling the spectral leakage between transient keyboard noise and vocal formants without destroying high-frequency clarity? And what’s your process for isolating whether a failure mode is the model’s training data gap or just an inference engine bottleneck?

Also curious about the practical side—when the DL filter starts producing those robotic, warbling artifacts, is it usually a sampling rate mismatch, insufficient training data for your specific voice profile, or an actual algorithmic limitation you can’t work around? I’m trying to figure out the root cause before I commit to one approach for tournaments where I can’t afford comms failing.