Microphone Noise Suppression DSP: Spectral Subtraction vs Deep Learning Filters: Advanced Implementation and Real-World Architecture

[Lucas Edwards]

I’ve been benchmarking RNNoise against a hand-rolled spectral subtraction pipeline for my streaming setup, and I’m torn on the latency tradeoff. Spectral subtraction sits around 3-5ms of latency, which is invisible for gameplay, but RNNoise and DeepFilterNet are pulling sub-20ms while doing a far better job on the click of my linear magnetic switches—yet I’m seeing occasional “underwater” artifacts during long sentences. For anyone running real-time inference on CPU vs GPU, what’s your actual latency budget before you feel the audio desync creeping in during ranked?

Mer importantly, how are you handling the spectral leakage when background noise shifts suddenly, like a fan spinning up mid-round? I’m using a noise profile updated every 10 seconds, but I suspect that’s too slow. Does anyone have experience with recursive estimation or Wiener filtering hybrids that adapt faster without introducing pumping artifacts? And for the deep learning route—have you quantized your models to INT8 for real deployment, and what’s the actual frame size you’re using (I’m currently on 10ms frames)? Would love concrete numbers on CPU utilization too, since I refuse to sacrifice FPS for cleaner voice chat.

"Daily Esport Driver:"

"• Ryzen 9 7950X3D / RTX 4070 Ti Super"

"• LG 27GR95QE 240Hz OLED"

"• Berlin, Germany"

[Author]

Posts