为电动汽车音质评估建立可复现的声学分析与机器学习基准
Automotive Sound Quality for EVs: Psychoacoustic Metrics with Reproducible AI/ML Baselines
- 整合国际标准听觉指标，构建轻量AI模型基线
- 在固定数据划分下实现90%以上分类准确率与高相关性
- 适合汽车声学研究、教学及电动车噪声优化应用
本文为电动汽车音质评估提供开放、可复现的参考框架，将标准化心理声学指标（响度、音调、粗糙度、波动强度）与轻量级机器学习基线相结合。采用ISO 532-1/2响度（sones）、DIN 45681音调及基于调制的描述符，并明确参数设定以确保可重复使用。建模方面，基于合成电动车场景，使用固定划分与种子训练逻辑回归、随机森林和SVM等简单模型，报告分类准确率与秩相关性作为端到端流程示例。整体音量以ITU-R BS.1770标准的LUFS衡量，心理声学分析则采用ISO-532响度。所有图表均通过代码脚本生成并锁定环境；代码与最小音频刺激已开源，支持教学、复现及扩展至逆变器啸叫等电动车特有噪声研究。
原文摘要 · Abstract (English)
We present an open, reproducible reference for automotive sound quality that connects standardized psychoacoustic metrics with lightweight AI/ML baselines, with a specific focus on electric vehicles (EVs). We implement loudness (ISO 532-1/2), tonality (DIN 45681), and modulation-based descriptors (roughness, fluctuation strength), and document assumptions and parameterizations for reliable reuse. For modeling, we provide simple, fully reproducible baselines (logistic regression, random forest, SVM) on synthetic EV-like cases using fixed splits and seeds, reporting accuracy and rank correlations as examples of end-to-end workflows rather than a comparative benchmark. Program-level normalization is reported in LUFS via ITU-R BS.1770, while psychoacoustic analysis uses ISO-532 loudness (sones). All figures and tables are regenerated by scripts with pinned environments; code and minimal audio stimuli are released under permissive licenses to support teaching, replication, and extension to EV-specific noise phenomena (e.g., inverter whine, reduced masking).
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。