Kolloquium - Details zum Vortrag
Sie werden über Vorträge rechtzeitig per E-Mail informiert werden, wenn Sie den Newsletter des kommunikationstechnischen Kolloquiums abonnieren.
Master-Vortrag: Deep Speech Inpainting for Masks Resulting from Noisy Spectra
Freitag, 9. Juli 2021
Speech enhancement belongs to the domain of noise detection, estimation and reductionwith the goal of increasing speech quality and intelligibility for speech signals subject toenvironmental noise. However, conventional methods typically insufficiently remove noisein challenging low-SNRnon-stationary noise environments.
In this thesis, we explore a twostage speech enhancement method utilizing the strengths of machine learning and deep speechinpainting for these difficult scenarios. Consisting out of a masking and inpainting stage,we deploy two powerful residualCNNs in an encoder-decoder structure, one for each stage.The purpose of the first stage is to mask a noisy-speech magnitude spectrogram using binarymasking such that it sufficiently removes noise but retains most of the important speechinformation. The second stage deals with reconstruction of the previously removedT-Fpointswith the goal of creating a corruption-free, continuous clean-speech signal. The results showthat our approach outperforms conventional speech enhancement methods and achieves similarstate-of-the-art performance regarding other speech inpainting studies. Compared to thenoisy-speech signal, our approach achieves up to 3 times higherPESQscores and substantialSNRimprovements of several decibels on ideal conditions. Furthermore, we demonstrate thatthe inpainting neural network can inpaint even highly masked speech if specifically trained for.
Alle Interessierten sind herzlich eingeladen, eine Anmeldung ist nicht erforderlich.