| Summary: | Speech enhancement for drone audition is made challenging by the strong ego-noise from the rotating motors and propellers, which leads to extremely low signal-to-noise ratios (e.g. SNR <inline-formula> <tex-math notation="LaTeX">$< -15$ </tex-math></inline-formula> dB) at onboard microphones. In this paper, we extensively assess the ability of single-channel deep learning approaches to ego-noise reduction on drones. We train twelve representative deep neural network (DNN) models, covering three operation domains (time-frequency magnitude domain, time-frequency complex domain and end-to-end time domain) and three distinct architectures (sequential, encoder-decoder and generative). We critically discuss and compare the performance of these models in extremely low-SNR scenarios, ranging from −30 to 0 dB. We show that time-frequency complex domain and UNet encoder-decoder architectures outperform other approaches on speech enhancement measures while providing a good trade-off with other criteria, such as model size, computation complexity and context length. The best-performing model is a UNet model operating in the time-frequency complex domain, which, at input SNR −15 dB, improves ESTOI from 0.1 to 0.4, PESQ from 1.0 to 1.9 and SI-SDR from −15 dB to 3.7 dB. Based on the insights drawn from these findings, we discuss future research in drone ego-noise reduction.
|
|---|