New Speech Models Target Real-Time Phone Scam Detection

Phone scams remain a persistent threat, and three new arXiv preprints approach the problem from the audio side, aiming to flag fraud during a live call rather than after the fact. The first paper, StreamFraudNet, focuses on weakly supervised incremental detection, where only conversation-level labels are available at training time and predictions must be updated in real time. This is a practical shift from offline analysis, as it targets the moment when intervention could still prevent harm.

The second paper, TeleAntiFraud 2.0, argues that benchmarks must keep pace with fraudsters, who constantly rewrite scripts to mimic routine service calls. It proposes a refreshable, profile-grounded, and audio-based evaluation framework, emphasizing that static datasets quickly become outdated. The third paper, FRAUDSkill, addresses a different bottleneck: large audio-language models can reason over speech but often fail to produce outputs that match a predefined label set. Its frozen-weight skill optimization method aims to align model outputs with structured fraud categories without fine-tuning the entire model.

All three papers share the premise that audio-based detection is viable and that real-world deployment demands more than raw accuracy—whether through incremental updates, fresh benchmarks, or constrained output formats. They differ in their primary focus: StreamFraudNet prioritizes timing, TeleAntiFraud 2.0 prioritizes data relevance, and FRAUDSkill prioritizes output structure. Together, they suggest that the field is maturing from proof-of-concept to deployment-oriented engineering, though none of the abstracts report end-to-end field results yet.