Santiago Pascual1, Joan Serrà2, Antonio Bonafonte1
1Universitat Politècnica de Catalunya, Barcelona, Spain
2Telefónica Research, Barcelona, Spain
This is a sample page of our work on Generative Adversarial Networks for End-to-End Speech Enhancement. In this work we build SEGAN and its improved version, SEGAN+. They prove to be competitive with our best log-power spectral DNN baseline configuration. In this page we show, for every utterance, the 6 systems: noisy (original contaminated signal), Wiener baseline, logMMSE baseline, DNN baseline with 7 context frames, SEGAN and SEGAN+.