overfitprolabse/subjective_audio_quality

这是一个大规模、平衡的音频质量感知评估数据集，包含612,020个示例，每个示例包含一对1秒的音频剪辑：高质量原始版本和经过各种音频编解码器处理的降质版本。每对剪辑都伴随一个感知质量分数（范围为0.0到1.0），由visqol-like算法生成。数据集的关键特征是其质量分数的平衡分布。与典型的音频数据集不同，这些数据集严重偏向高质量的示例，而这个数据集已被精心策划，以确保在整个质量范围内均匀分布，从“糟糕”（MOS ≈ 1.0）到“优秀”（MOS ≈ 5.0）。这使得它非常适合训练能够在所有质量级别上可靠运行的鲁棒回归模型，尤其是在低质量范围内。数据集以`webdataset`格式提供，适合高性能、顺序I/O，非常适合大规模深度学习工作负载。

This is a large-scale, balanced dataset designed for training models for perceptual audio quality assessment. It consists of **612,020** examples, each containing a pair of 1-second audio clips: a high-quality original and a degraded version processed by various audio codecs. Each pair is accompanied by a perceptual quality score **(ranging from 0.0 to 1.0)** generated by visqol-like algorithms. The key feature of this dataset is its **balanced distribution of quality scores**. Unlike typical audio datasets, which are heavily skewed towards high-quality examples, this dataset has been carefully curated to ensure a uniform distribution across the entire quality spectrum, from "terrible" (MOS ≈ 1.0) to "excellent" (MOS ≈ 5.0). This makes it exceptionally well-suited for training robust regression models that perform reliably across all quality levels, especially in the challenging low-quality range. The dataset is provided in the `webdataset` format for high-performance, sequential I/O, making it ideal for large-scale deep learning workloads.