Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This work introduces AEROMambaP, an efficient audio enhancement architecture that replaces traditional attention and LSTM layers with a Mamba state-space model and a novel differentiable perceptual loss derived from the perceptual audio quality measure (PAQM). The architecture significantly reduces GPU memory usage during training by 2 to 4 times and achieves a 14× speedup during inference while maintaining high audio quality. In subjective tests, AEROMambaP outperforms its predecessor AERO by 15% in perceived quality, and its variant AEROMambaPS̄ enhances MP3-encoded audio at 32 kbps, achieving a 52% higher quality rating compared to AEROMambaP.
AEROMambaP achieves a remarkable 14× inference speedup while enhancing audio quality, setting a new standard for efficient audio enhancement.
Audio enhancement aims to improve the perceived quality of audio signals. Initially, to address bandwidth extension, this work proposes AEROMambaP, an efficient variant of the AERO super-resolution architecture in which attention and long short-term memory layers are replaced by the Mamba state-space model, together with a newly developed differentiable perceptual loss derived from the perceptual audio quality measure (PAQM). During training, the architecture requires two to fours times less GPU memory than the baseline; during inference, it achieves a 14× speedup while using one-fifth of the GPU memory. When upsampling both a piano dataset and MUSDB18 from 11.025 kHz to 44.1 kHz, subjective listening tests show that AEROMambaP outperforms AERO by 15% in perceived quality scores. Next, to enhance of audio signals highly compressed by lossy coding, it is further proposed AEROMambaPS̄, which applies the same framework but replaces short-time Fourier transform reconstruction losses with the PAQM loss, specifically to enhance MP3-encoded audio at 32 kbps. In listening evaluations, AEROMambaPS̄ achieves 52% higher quality rating than AEROMambaP when restoring compressed audio. These results demonstrate that PAQM-driven training coupled with lightweight state-space modeling yields high perceptual quality and computational efficiency in both band-limited and compressed audio scenarios.