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.

Bachelor-Vortrag: Deep Learning-Based Speech Synthesis as Post-Processing of a Noise Reduction

Konstantin Wehmeyer
Mittwoch, 18. Januar 2023
14:00 Uhr

Audio and speech signals are often disturbed by noise signals in frequency- and/or time-limited parts. To attenuate or remove these distortions, several methods, including deep learning- based approaches, are known. Often, however, only the magnitude spectrum is processed and the phase spectrum is taken over unchanged due to its comparatively lower relevance. Consequently, the noisy phase is reused when synthesizing the waveform from the processed magnitude spectrum. Therefore, distortions in the magnitude spectrum can be reduced, but not in the phase spectrum which inevitably leads to a deterioration in speech quality and intelligibility.

This thesis presents methods that allow a reconstruction of the phase spectrum of speech signals based on noise-reduced magnitude spectra. At the Institute of Communication Systems at RWTH Aachen University a phase reconstruction algorithm was developed and this algorithm has already been evaluated in a previous study for the case of smoothed magnitude spectra. It was shown that the deep neural network (DNN) used can benefit from targeted training on the smoothed magnitude spectra even without further modification of the network structures. However, even slight smearing of the magnitude spectra already leads to a significant loss in performance compared to the use of perfect magnitude spectra. In this work, therefore, the DNNs used are optimized for the case of noise-reduced magnitude spectra.

Several deep learning-based models are introduced and compared with each other and with the models already developed. Their properties and aspects such as causality are addressed. Moreover, a new loss function and assessment measure specifically designed to estimate and assess the phase spectrum of speech signals is developed and tested. In order to be able to evaluate the results as independently as possible of a specific type of noise reduction, ideal masks are developed, used, and discussed.