Kolloquium - Details zum Vortrag
Sie werden über Vorträge rechtzeitig per E-Mail informiert werden, wenn Sie den Newsletter des kommunikationstechnischen Kolloquiums abonnieren.
Master-Vortrag: Speech Inpainting Using Recurrent Neural Networks
Mittwoch, 20. Januar 2021
Speech signals are often subject to transient noise signals that can appear anywhere in the frequency spectrum. When the noise overlaps with frequencies of human speech, the removal of the noise inevitably leads to a deterioration in the quality of speech and a loss of information. Because speech is highly predictable in time, a speech inpainting neural network is presented which includes recurrent neural networks in the form of LSTMs whose strength lies in the prediction of sequential data. Furthermore, the SHAP algorithm is employed to gain an understanding of the impact of individual input features on the output and the network’s dependence on individual time steps in the input sequences.
The results demonstrate that LSTMs are well-suited to solve the problem of speech inpainting, outperforming other networks based on fully-connected layers only. An increase in the PESQ and LSD score is observed compared to the corrupted signal. The increase is further enhanced when the sequences include not only past but also future data samples. Interpretation of the trained networks concludes that the prediction of output features is primarily based on their counterpart in the input features with the largest emphasis in the sequences being put on the current time step.
Alle Interessierten sind herzlich eingeladen, eine Anmeldung ist nicht erforderlich.