yairamr/SE-Probe-data

SE-Probe完整结果数据集是一个用于语音增强和可解释性研究的预计算数据集，包含CKA（中心核对齐）、扩散映射和探测结果。该数据集支持论文《Where Does Speech Enhancement Adapt? Probing Study Under Controlled Degradation》（Amar, Ivry, Cohen, 2026）的研究，提供了针对不同架构（如MUSE、MP-SENet、Demucs）在加性噪声信噪比和DEMAND噪声类型下的CKA值，以及每行音频质量指标（如PESQ、STOI、SI-SDR、DNSMOS系列）。此外，还包括每epoch的混响微调MUSE探测结果和预计算的扩散映射嵌入，总大小约3.1 GB。SE-Probe full results dataset is a pre-computed dataset containing CKA, diffusion-map, and probing results for the SE-Probe project, supporting the research paper Where Does Speech Enhancement Adapt? Probing Study Under Controlled Degradation (Amar, Ivry, Cohen, 2026). It includes per-architecture CKA values across additive-noise SNRs and DEMAND noise types with audio quality metrics (PESQ, STOI, SI-SDR, DNSMOS family), per-epoch reverb fine-tuned MUSE probing results, and pre-computed diffusion-map embeddings per model, SNR, noise, and layer, totaling approximately 3.1 GB.