Indoor context perception relies on massive IoT deployments, where Wi‑Fi Channel State Information (CSI) can passively capture human activities and environmental dynamics. Unlike vision or audio, CSI lacks natural textual annotations (captions, transcripts), making it hard to plug into language‑centric multimodal systems. To bridge this gap we introduce CSI-TextBench, a large‑scale dataset that unifies three existing sensing benchmarks into 285,790 samples across seven tasks. We generate 2,291 hierarchical, label‑derived natural‑language descriptions and create about 2.0 million CSI‑text training tuples via many‑to‑many assignment. The benchmark supports fine‑grained tasks such as activity recognition, proximity estimation, and identity verification under a single interface. We evaluate five alignment methods from four paradigm families (contrastive learning, cross‑modal mapping, generative alignment, label‑based mapping). Results show that contrastive CSI‑text alignment matches supervised classifiers without using class labels, adapts quickly to unseen classes with few examples, and responds to open‑vocabulary natural‑language prompts. Moreover, a single unified model enables compositional scene understanding, answering joint queries over activity, proximity, and identity, achieving 86.9% top‑1 accuracy on 140 composite scenes—outperforming ensembles of task‑specific classifiers. CSI-TextBench thus makes ambient sensing a language‑indexed perception layer, allowing indoor contextual awareness to be directly leveraged by language‑driven AI systems.
Review