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 for Adaptive Crosstalk Cancellation Based on Linear Dynamical Models

Kang Pei
Donnerstag, 15. April 2021
14:30 Uhr
virtueller Konferenzraum

Playback of binaural recordings is quite popular not only in academia but also in commercial and entertainment industry. Compared with playback based on headphones, the implementation of that by loudspeakers is more challenging because of the "Crosstalk" phenomenon: the sound signals from left or right loudspeaker will both affect the sound perception on both ears and, as a result, important sonic cues (Inter-aural Time Difference and Inter-aural Level Difference) will be influenced and the 3D sound effect is hard to perceive. To conquer such "Crosstalk" phenomenon, carefully designed "Crosstalk Cancellation filter" (CTC filter) is utilized to pre-process binaural signals.

The calculation of CTC filter is based on "Head Related Impulse Responses" (HRIRs) estimation. Kalman filter, as one of the most popular method in system idenfitication, is widely utilized in HRIRs estimation. Researches show that the current HRIRs estimator based on Kalman filter could generate CTC filter that attenuates the crosstalk of binaural sounds. However, these researches are based on the conditions either the listener sits still or has a constant rotational velocity. In this thesis, we will design a novel architecture to estimate HRIRs for CTC filter when the rotational velocity of the listener is time-variant. The innovative part of such architecture is that based in Expectation-Maximation (EM) algorithm designed for "Linear Dynamic System" (LDS) parameters estimation, different sets of LDS parameters are pretrained and will be selected for HRIRs estimation under certain rotational velocity. Since EM algorithm is an iterative algorithm, initial parameters should be selected with caution. In this thesis, we first validate the performance of our novel architecture based on single-channel base where the HRIR sequence has Markov property and is affected by Gaussian noises. Then, simulation experiments, also based on single-channel, with HRIR sequence generated from free-field propagation are implemented and analyzed. The results show that HRIR could be well estimated in single-channel case with such novel architecture and we need further investigation to discover its performance on attenuating crosstalk in multi-channel case.

zurück