Microphone Noise Suppression DSP: Spectral Subtraction vs Deep Learning Filters: Fundamental Mechanics and Deep Mathematical Modeling

[Grace Wright]

I’ve been benchmarking spectral subtraction against learned filters for my streaming setup, and I want to understand where each actually breaks down under real competitive conditions. For spectral subtraction, how are you handling the musical noise artifact problem when the over-subtraction factor alpha gets pushed for heavier keyboard noise? I’ve read that the phase estimation gets neglected entirely in the classic Wiener-filter derivation, but in fast-talking team comms, isn’t that phase distortion the real killer for intelligibility? And for the deep learning approaches—how are you dealing with the inference latency budget? I’m seeing 8-15ms added round-trip on GPU-accelerated models, which is noticeable in ranked where every millisecond of audio feedback matters.

Specifically: when you’re eliminating hot mechanical keyboard clicks (those sharp transients around 2-4kHz), are you training your model on domain-specific noise or relying on generalization? I’ve noticed models trained on noisy speech datasets completely miss the impulsive characteristics of key switches. Has anyone done a rigorous ablation comparing SNR improvement versus perceived voice naturalness, or is the community mostly going by raw dB numbers? And what’s your practical ceiling for acceptable latency before comms start feeling laggy?

"NVIDIA Reflex + G-Sync + V-Sync ON + Frame Cap (-3 FPS) = 0 Tearing, Lowest Latency"

[Grace Gray]

Great topic to crack open. Spectral subtraction is the workhorse here because it’s cheap and explainable, but the math reveals why it still leaks artifacts. You’re essentially working in the STFT domain, estimating a noise profile during silence gates (VAD-triggered), then subtracting it frame-by-frame: |Ŝ(ω)|² = max(|X(ω)|² − β|N(

"☕ Double shot espresso + tenkeyless custom mechanical board + dim ambient light."

[Author]

Posts