New I²SRF-TFCKD Framework Boosts Speech Enhancement with 2x Distillation Efficiency

In a new paper, a team led by Jiaming Cheng (including researchers from China and Germany) introduces I²SRF-TFCKD — a distillation framework that dramatically improves speech enhancement in low-complexity models. The approach addresses a key limitation of prior knowledge distillation (KD) methods, which treat all frequency and time regions equally. Instead, I²SRF-TFCKD explicitly leverages time-frequency differential information through two novel mechanisms: (1) intra-set and inter-set recursive fusion that pairs multi-layer teacher-student features within correlated sets, then fuses representative features across sets for global knowledge circulation; (2) a multi-layer interactive distillation based on dual-stream time-frequency cross-calibration, which computes similarity weights in the time and frequency domains separately and applies cross-weighting to allocate distillation contributions per layer according to speech characteristics.

The proposed framework is applied to the DPDCRN architecture, which won the speech enhancement track of the L3DAS23 challenge. Experiments on both single-channel and multi-channel speech enhancement datasets show that the student model trained with I²SRF-TFCKD consistently outperforms students trained with other KD schemes — even surpassing some larger teacher models in certain objective metrics. The method is particularly effective at preserving speech intelligibility while suppressing noise, thanks to the calibrated focus on informative time-frequency bins. This work has been submitted to Neural Networks and is available on arXiv (2506.13127). For practitioners, it means high-quality speech enhancement can now be achieved with significantly smaller, faster models — ideal for edge devices and real-time applications.