Filamentary Convolution for SLI: A Brain-Inspired Approach with High Efficiency

• DOI: 10.3390/s25103085
• Publication Date: 2025-05-13T15:31:49Z
• Authors: Boyuan Zhang; Xibang Yang; Tong Xie; Shuyuan Zhu; Bing Zeng
• Keywords: Convolution (computer science)

Spoken language identification (SLI) relies on detecting key frequency characteristics like pitch, tone, and rhythm. While the short-time Fourier transform (STFT) generates time–frequency acoustic features (TFAF) for deep learning networks (DLNs), rectangular convolution kernels cause mixing aliasing, degrading feature extraction. We propose filamentary to replace kernels, reducing parameters while preserving inter-frame by focusing solely patterns. Visualization confirms its enhanced sensitivity critical variations (e.g., intonation, rhythm) recognition. Evaluated via self-built datasets cross-validated with public corpora, improves low-level extraction efficiency synergizes temporal models (LSTM/TDNN) boost This method addresses aliasing limitations maintaining computational in SLI systems.