SMoLnet-T: An Efficient Complex-spectral Mapping Speech Enhancement Approach with Frame-wise CNN and Spectral Combination Transformer for Drone Audition

Speech enhancement for drone audition applications is challenging due to the low SNR with large spectra feature overlap and limited computing resources. We propose SMoLnet-T, a complex spectral mapping approach with frame-wise CNN and newly-formulated spectral combination transformers. SMoLnet-T incorporates dilated CNN to extract spectral maps of high-frequency resolution for its transformers. This allows it to focus on a higher level of abstraction and determine the combination of spectral maps is crucial for enhancement across a large temporal context. Experiment results with noise recorded from a hovering drone highlight the efficacy of SMoLnet-T over DPTNet with significantly lower computational requirements and speech distortion while achieving improved speech intelligibility under SNR < −23 dB.