On the Impact of Frequency Resolution on Female and Male Speech in DNN-Based Noise Reduction Systems

Neural network-based noise reduction has shown huge potential and superior performance in comparison to traditional statistical signal processing methods. In low-power devices such as hearing aids, however, constraints such as low complexity and algorithmic latency, as well as trade-offs between frequency resolution and algorithmic latency, must be considered. In this study, we vary the frequency resolution of an established low-complexity DNN-based algorithm while keeping the number of network parameters as well as the time resolution fixed. Furthermore, we compare the results to a traditional Wiener filter-based algorithm. We find that the performance of DNN-based noise reduction algorithms is contingent upon the frequency resolution and the fundamental frequency of the speech signal, in a manner analogous to that observed with traditional Wiener filters. When evaluating the performance dependent on the fundamental frequency of the speech signal, we find that the DNN-based noise reduction in female speech benefits more from higher frequency resolution than noise reduction in male speech.