This site acts as supplementary material to our work on Directivity-Conditioned Low-Latency Neural Filtering for Speech Enhancement in Hearing Aids. We present a low-latency solution for neural directional filters for a hearing-device microphone setup that achieves comparable results to a baseline without low-latency constraints.

Synthetic dataset

We create a synthetic dataset to enable training neural directional filtering approaches on a hearing device microphone array setup. The dataset consists of five-speaker mixtures in a reveberant room, which are spatialized using the Hearing Aid Related Transfer Functions from . One individual head and one kemar head is used to evaluate the models. The reverberation time varies between 0.2-0.5s.

Directivity-Conditioned Low-Latency Neural Filtering for Speech Enhancement in Hearing Aids

Lennart Uphaus, André Merboldt, Markus Hofbauer, Timo Gerkmann

Select an audio file:

Audio
Noisy:	PESQ ESTOI SISDR 1.23 0.60 -3.76
FiLM-JNF-32ms :	PESQ ESTOI SISDR 2.72 0.83 6.53
FiLM-JNF-8ms:	PESQ ESTOI SISDR 1.77 0.73 2.48
FiLM-OSN-8ms-L1:	PESQ ESTOI SISDR 2.48 0.83 4.64
FiLM-OSN-8ms-L1+IPD:	PESQ ESTOI SISDR 2.56 0.82 4.64
Clean: