Compressed Representation of Cepstral Coefficients via Recurrent Neural Networks for Informed Speech Enhancement.

Speech enhancement is one of the biggest challenges in hearing prosthetics. In face-to-face communication devices have to estimate the signal of interest, but playback of speech signals from an electronic device opens up new opportunities. Audio signals can be enriched with hidden data, which can subsequently be decoded by the receiver. We investigate a hybrid strategy made of signal processing and RNN (Recurrent Neural Networks) to calculate and compress cepstral coefficients: these are descriptors of the speech signal, which can be embedded in the signal itself and used at the receiver’s end to perform an Informed Speech Enhancement. Objective evaluations showed an increase in speech quality for noisy signals enhanced with our method.