Colloquium - Details

Master-Presentation: Low-Complexity Noise Reduction for Hearables

Eylül Ercandogu
Monday, November 25, 2024
14:00 PM
IKS 4G | zoom

In everyday life, noise interference significantly affects speech quality and intelligibility in various applications, such as hearing aids, earbuds, and mobile communication devices. These devices, known for their compactness and limited processing power, face challenges in implementing complex signal-processing algorithms.

This work focuses on designing and evaluating efficient, real-time speech enhancement models tailored for low-complexity hearables. We adopted the Grouped Temporal Convolutional Recurrent Network (GTCRN), which reduces computational demands through efficient design elements. The real-time adaptation, RT-GTCRN, achieves competitive performance with minor quality trade-offs, making it well-suited for deployment on low-power devices. Additionally, we explored the axial self-attention mechanism to capture long-range dependencies efficiently. To improve speech quality, we incorporated advanced loss functions, including perceptual losses, which led to significant improvements in perceptual speech quality metrics. To enhance model generalization, we investigated data augmentation techniques and evaluated their impact across inter- and intra-corpus assessments. These data augmentation methods demonstrated promising improvements in handling diverse acoustic scenarios. Furthermore, knowledge distillation techniques were explored to enhance model efficiency by maintaining a small model while improving its performance. While the method demonstrated promise, further optimization could be beneficial to bridge the performance gap between the teacher and student models.

The RT-GTCRN model was successfully implemented on the GAP9 processor, yielding promising results despite the observation of minor artifacts during implementation. These findings highlight potential areas for future refinement.

back