Kolloquium - Details zum Vortrag
Sie werden über Vorträge rechtzeitig per E-Mail informiert werden, wenn Sie den Newsletter des kommunikationstechnischen Kolloquiums abonnieren.
Master-Vortrag: Enhanced Own Voice Perception for Hearables Using Machine Learning
Mittwoch, 23. Juni 2021
Modern headphones implement sophisticated signal processing and control algorithms to improve the user’s hearing and listening experience. Especially new generation in-ear headphones, so-called hearables, integrate these technologies, such as active noise control, on a large scale. A remaining issue with hearables is the sound degradation introduced through sealing the ear canal. This degradation is called the Occlusion Effect. It is often described as boomy and hollow perception of the speaker’s own voice. While Active Noise Cancellation reduces ambient noise, the Occlusion Effect still persists. In order to improve the speaker’s own voice perception and alleviate the Occlusion Effect, it is required to play back some of the voice signal without amplifying ambient noise.
In this thesis a novel approach for selective hear-through based on deep neural networks is investigated. Integrated into a feed-forward ANC topology the proposed neural network utilizes ambient and in-ear microphone signals for an estimation of the speaker’s own voice. Speech samples for both microphone signals were simulated through augmentation of the DARPA TIMIT speech corpus. This included virtually occluding the samples for one channel and adding real-world noise samples to the other. The proposed network was found to reduce ambient noise from the test dataset to a barely audible extent. Metrics for quality and intelligibility suggest superior denoising performance in comparison with one channel network benchmarks. Despite the relatively small training dataset, the network may also regularize well when trained with measurement data.
Alle Interessierten sind herzlich eingeladen, eine Anmeldung ist nicht erforderlich.