Kolloquium - Details zum Vortrag
Sie werden über Vorträge rechtzeitig per E-Mail informiert werden, wenn Sie den Newsletter des kommunikationstechnischen Kolloquiums abonnieren.
Alle Interessierten sind herzlich eingeladen, eine Anmeldung ist nicht erforderlich.
Master-Vortrag: Machine Learning for Primary-Ambient Extraction of Audio Signals
Faezeh Abediostad
Donnerstag, 24. August 2017
11:00 Uhr
Hörsaal 4G
Extraction of primary and ambient components from a stereo audio signal is a crucial preprocessing stage for many audio rendering tasks including stereo to multi-channel up-mixing. A study on the functionality of machine learning approaches for solving this problem is performed. While a fully supervised learning method requires expensive computations of hand-coded dictionaries of distinctive characteristics of ambient and primary audio, the generality of the audio scene makes this task almost impossible. Our results indicate that through unsupervised learning, stereo information alone can be organized to reveal a pattern for primarity. Such developed higher-level abstractions of channel dependencies can be further utilized in form of a probabilistic time-frequency mask through supervised discriminative fine-tuning. Consequently, the presented multi-layer neural network performs a soft-margin classification of STFT representations into directed and diffused parts
