clatter-1/XF-Denoise

Far-field speech enhancement plays a crucial role in speech signal processing, primarily aimed at improving speech intelligibility and other speech-based applications in real-world scenarios. However, models trained on simulated data or existing real-world far-field datasets often exhibit limited performance in extreme far-field conditions. To address these challenges, we present XF-Denoise, a real-world dataset specifically designed for extreme far-field speech enhancement. The dataset comprises 70 hours of paired near-field and far-field speech recordings collected across 10 diverse scenarios, with far-field distances extending up to 8 meters. The inclusion of 8-meter recordings, along with diverse indoor and outdoor scenarios, provides valuable data for developing and evaluating speech enhancement models in various far-field conditions, offering significant advantages for real-world applications. Experiments on mainstream discriminative and diffusion models demonstrate that models trained on the XF-Denoise dataset achieve better performance in real far-field environments, validating the dataset's effectiveness and practical utility.