FFT-ConvAE is a lightweight speech-enhancement method that formulates audio reconstruction as an inverse (deconvolution) problem and solves it by learning only the Fourier magnitudes of degraded signals. The full-signal FFT feeds a convolutional autoencoder with purely linear activations; the degraded signal's phase is reused at reconstruction because phase training overfits. Achieving real-time factors far below the Helsinki Speech Challenge 2024 cap, it took second place, performing strongly on filtering degradations while struggling on reverberation, whose deconvolution is more ill-posed — a boundary its appendix explains through the singular-value decay that separates stably recoverable features from unstable ones.
FFT-ConvAE is a lightweight speech-enhancement model that combines the discrete Fourier transform with a convolutional autoencoder, developed for the Helsinki Speech Challenge 2024 (HSC2024), where i…