Slowdio: Audio Time-Scaling for Slow Motion Sports Videos

Today, most of sports television broadcasts feature slow motion playbacks. Until now, these have been silent. In this thesis, we present several state of the art methods for time-scaling of audio signals and study their behavior when applied to a new database of sports recordings. We argue that the underlying models used to develop these methods do not correspond to the noisy audio signals recorded during sports events. Besides, transient sounds need to be detected and processed separately from the rest of the signal, which proves difficult in the typical noisy environments of sports events. Based on hypotheses that better fit the actual content of these recordings, we develop a new method that produces convincing time-scaled audio signals while implicitly handling transients. Furthermore, we introduce a new time-scaling approach for harmonic sounds such as speech and single-instrument music recordings.