Machine Learning based Evaluation of Feedback Cancellation Systems in Hearing Aids
Author
Lauridsen, Jakob Sloth
Term
4. term
Education
Publication year
2020
Submitted on
2020-06-04
Pages
53
Abstract
This thesis explores the use of machine learning to evaluate feedback cancellation in hearing aids by automatically detecting artifacts such as howling and the so-called alpha artifact that can occur when adaptive feedback cancellation (AFC) and spectro-temporal modulation (STM) are engaged. The acoustic feedback problem is framed via closed-loop stability (Nyquist) and a review of state-of-the-art AFC and its side effects. To replace time-consuming manual annotation, artifact detection is posed as a supervised classification task using spectro-temporal representations (STFT) from hearing-aid signals together with a reference channel. Single-hidden-layer feedforward neural networks with 1024 hidden units are trained and evaluated with standard metrics, and performance is compared to an existing MATLAB-based detection system from Oticon. Several ways of processing spectral inputs and the reference are investigated, including omitting the reference, using absolute differences, and correlation between channels. Findings indicate that a single-layer network can achieve performance comparable to the existing detector, and that similar performance is possible without the reference channel. This increases flexibility in test environments and suggests potential for additional uses, such as online evaluation of feedback cancellation systems.
Dette speciale undersøger, hvordan maskinlæring kan bruges til at evaluere feedback-annulleringssystemer i høreapparater ved automatisk at opdage artefakter som hyl (howling) og den såkaldte alpha-artefakt, der kan opstå, når adaptive feedbackmetoder og spectro-temporal modulation (STM) anvendes. Problemet med akustisk feedback beskrives via lukket kredsløbsstabilitet (Nyquist) og en gennemgang af state-of-the-art adaptive feedback-annullering (AFC) og dens bivirkninger. For at erstatte tidskrævende manuel annotering formuleres artefektdetektion som en klassifikationsopgave baseret på spektro-temporale repræsentationer (STFT) fra høreapparatets signaler samt en referencekanal. Enkeltlags feedforward-neurale netværk med 1024 skjulte neuroner trænes og evalueres med standardmål, og præstationen sammenlignes med et eksisterende MATLAB-baseret detektionssystem fra Oticon. Forskellige måder at behandle spektrale data og referencekanalen på undersøges, herunder at udelade referencen, bruge absolut forskel og korrelation mellem kanalerne. Resultaterne viser, at et enkeltlags netværk kan opnå en ydeevne på niveau med den eksisterende løsning, og at tilsvarende ydeevne kan opnås uden referencekanalen. Det øger fleksibiliteten i testmiljøer og peger på mulige ekstra anvendelser, såsom online evaluering af feedback-annulleringssystemer.
[This apstract has been generated with the help of AI directly from the project full text]
Keywords