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 Techniques to Reconstruct Lost Parts of Speech Signals
Marcel Czaplinski
Donnerstag, 6. Juni 2019
11:15 Uhr
Hörsaal 4G
Applications for speech transmission and mobile communication have high demands for speech intelligibility and authenticity. Errors and corruptions of different types are commonly occuring. Often, parts of the speech signal are missing completely.
A major task in speech processing and transmission is the enhancement of a speech signal and minimizing distortions that occur as a result of corruptions. If the distorted signal parts of the speech signal cannot be restored, they can be dropped completely and reconstructed to a certain degree from the uncorrupted parts of the signal. This technique is the core idea of the well-known packet loss concealment (PLC) and bandwidth extension (BWE). However, these tools assume the missing parts to be of time or frequency limited shape.
Speech inpainting, the task of the reconstruction of lost parts of a speech signal of any shape, extends BWE and PLC to a generalized concept. Some dictionary based speech inpainters have been proposed from various researchers in the past. Despite the progress and promising results of recent machine learning research from related topics like image inpainting, not many endeavours have been made to use signal processing and machine learning jointly to build speech inpainters.
A general framework and overview of a machine learning assisted speech inpainter will be provided. A selection of preprocessing tools and algorithms will be analyzed in the context of different corruption types and the results compared to a simple interpolation algorithm. Furthermore, time and frequency interpolation capabilities of algorithms and speech features will become interpretable and new insights into existing problems will be granted.
