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: Konzepte zur Audiosignal-Analyse und -Synthese mit zeitvariabler Filterung
Mathew Kavalekalam
29. Juli 2014
11:30 Uhr
Hörsaal 4G IKS
In this thesis, a time domain noise reduction technique based on the Wiener filter and autoregressive modeling of speech signals has been proposed. Two fundamental issues to be tackled in this thesis are the estimation and the adaptation of the filter coefficients.
The assumption that both the clean speech signal and the noisy speech signal can be modeled as an autoregressive (AR) process was exploited for the derivation of the filter structure. Two different methods for estimating the parameters of the autoregressive model were investigated. The first method uses the linear prediction coefficients (LPC) of the speech signals as the parameters of the AR model. A limitation of the LPC based AR model is that the LPC based spectral envelope does not model the medium and high pitch voiced speech very well. To overcome this limitation an alternative method for estimating the parameters of the all pole model was carried out which is based on the Minimum Variance Distortionless Response (MVDR) spectrum. As the clean speech linear prediction coefficients are required for estimating the filter coefficients in the recursive part of the filter structure, these LPC must be estimated from the noisy speech signal.In this thesis we have investigated different methods for the estimation of the clean speech LPC. The estimation can be supported by a priori information about speech and noise spectral shapes stored in pre-trained codebooks.
Four different types of filter adaptation methods were investigated. Besides an optimal recursive smoothing parameter for the update was derived based on the minimization of a cost function which depends on the frame size.
In this thesis it could be shown that time domain filtering techniques has higher performance than conventional frequency domain enhancement methods in terms of objective measures. It was also observed that time domain noise reduction techniques are robust to musical noise which often occurs during conventional noise reduction approaches.
