aTENNuate: Optimized Real-time Speech Enhancement with Deep SSMs on Raw Audio

We present aTENNuate, a simple deep state-space autoencoder configured forefficient online raw speech enhancement in an end-to-end fashion. The network'sperformance is primarily evaluated on raw speech denoising, with additionalassessments on tasks such as super-resolution and de-quantization. We benchmarkaTENNuate on the VoiceBank + DEMAND and the Microsoft DNS1 synthetic test sets.The network outperforms previous real-time denoising models in terms of PESQscore, parameter count, MACs, and latency. Even as a raw waveform processingmodel, the model maintains high fidelity to the clean signal with minimalaudible artifacts. In addition, the model remains performant even when thenoisy input is compressed down to 4000Hz and 4 bits, suggesting general speechenhancement capabilities in low-resource environments. Try it out by pipinstall attenuate