Using recurrences in time and frequency within U-net architecture for speech enhancement
This work addresses speech enhancement for noisy audio processing, presenting an incremental improvement over existing U-net architectures.
The paper tackled the trade-off between receptive field size, parameters, and spatial resolution in fully-convolutional neural networks for speech enhancement by proposing a novel design combining convolutional and recurrent layers, resulting in clear advantages in SDR, SIR, and STOI metrics over state-of-the-art models on TIMIT and NOISEX-92 data.
When designing fully-convolutional neural network, there is a trade-off between receptive field size, number of parameters and spatial resolution of features in deeper layers of the network. In this work we present a novel network design based on combination of many convolutional and recurrent layers that solves these dilemmas. We compare our solution with U-nets based models known from the literature and other baseline models on speech enhancement task. We test our solution on TIMIT speech utterances combined with noise segments extracted from NOISEX-92 database and show clear advantage of proposed solution in terms of SDR (signal-to-distortion ratio), SIR (signal-to-interference ratio) and STOI (spectro-temporal objective intelligibility) metrics compared to the current state-of-the-art.