Leverage Sub-Band and Cross-Band Information in Real-Time Speech Signal Improvement

Le résumé fourni par la source

Sub-band and cross-band information have proven beneficial in improving high-resolution characteristics for speech separation and enhancement tasks. Nevertheless, the extraction of high-resolution features frequently results in significant computing complexity, which contradicts the low-complexity demands of real-time systems. To address this difficulty, conventional methodologies often employ an autoencoder design that compresses frequency data to diminish computing demands and encapsulates global information within the bottleneck. Nonetheless, these models frequently neglect the significance of sub-band and cross-band parameters in speech signal processing. To overcome this issue, we offer a real-time model that efficiently utilizes both sub-band and cross-band information at high resolution, while adhering to the autoencoder architecture to meet real-time requirements. The suggested model demonstrates exceptional performance across multiple speech quality metrics, attaining a SIG score of 3.49 and an OVRL score of 3.01 on the blind test set of the ICASSP 2024 Speech Signal Improvement Challenge, with a real-time factor (RTF) of 0.30.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.