Jan Rennies(Fraunhofer IDMT, Hearing, Speech and Audio Technology), Henning Schepker(University of Oldenburg, Signal Processing Group, Oldenburg), Cassia Valentini-Botinhao(The Centre for Speech Technology Research, University of Edinburgh) and Martin Cooke(Basque Foundation for Science, Bilbao)
Understanding speech played back in noisy and reverberant conditions remains a challenging task. This paper describes the Hurricane Challenge 2.0, the second large-scale evaluation of algorithms aiming to solve the near-end listening enhancement problem. The challenge consisted of modifying German, English, and Spanish speech, which was then evaluated by a total of 187 listeners at three sites. Nine algorithms participated in the challenge. Results indicate a large variability in performance between the algorithms, and that some entries achieved large speech intelligibility benefits. The largest observed benefits corresponded to intensity changes of about 7 dB, which exceeded the results obtained in the previous challenge despite more complex listening conditions. A priori information about the acoustic conditions did not provide a general advantage.