Microphone Noise Suppression DSP: Spectral Subtraction vs Deep Learning Filters: Historical Paradigm Evolution and Next-Gen Architectural Projections

[Liam Watson]

Real talk: I’ve spent the last few months benching spectral subtraction algorithms in my custom comms rig because they’re eating my high-frequency voice detail every time I try to kill the mechanical switch noise on my board. Those old Wiener-filter approaches were the standard for years, but the artifacts they leave—mostly that underwater, muffled “robotic” sound during team fights—have gotten unbearable for competitive play where every syllable of callout matters. So my first question is genuine: at what point did the community collectively move away from classical spectral subtraction, and what specific deep learning filters (RNNoise, NVIDIA Broadcast’s model, or custom Whisper-based pipelines) actually preserved voice intelligibility while still nailing the click and thock of a hot-gswap board?

Second, the latency argument is everything for me. Spectral subtraction runs near-zero-latency, but DL models add 20-50ms of processing delay that I can literally feel in ranked. Where’s the practical sweet spot now? Are there quantized or ONNX-optimized models that keep sub-15ms latency without reverting to legacy artifacts? And what architectural direction are you all betting on—self-supervised models likewav2vec, or are transformer-based noise estimators the real future for edge deployment?

"Benchmark results tested on clean Windows 11 with Core Isolation disabled."

[Daniel Torres]

Spectral subtraction has been the backbone of real-time noise suppression for decades, and honestly it’s still the pragmatic choice for competitive play where latency is non-negotiable. The classic approach—estimating the noise floor during pauses via minimum-statistics tracking and subtracting it from each FFT bin—suffers from that telltale “musical noise” where random phase residuals create those eerie tonal blips across the spectrum. Over-subpression factors and the shift toward minimum mean-square error (MMSE) estimators using a priori/a posteriori SNR ratios mitigated this significantly, and Wiener-style

"Competitive PC enthusiast."

[Author]

Posts