USC's SPARC benchmark shows transcripts still leak speaker identity to LLMs
USC · hf · 2026-10-20
USC researchers introduce SPARC (Spoken Profile Authorship Recognition Corpus), probing a blind spot in speech anonymization: even after acoustic features are stripped, transcripts retain word choice, syntax and discourse habits. Built from two podcast corpora with 12,500+ speakers and 1.5M quality-filtered utterances, SPARC benchmarks pairwise speaker and demographic-attribute verification from text. LLMs can exploit these linguistic traces to link anonymous transcripts to known speakers, while releasing only brief text excerpts reduces linkage risk.
More from Safety
- NIST, NSA and CISA Publish 10 Key Practices for Incident Response in Superintelligence Training — luisdans · 2026-10-20
- DeepMind's Demis Hassabis Warns of AI-Powered Hacks at Dublin Summit — BusyNomad · 2026-10-20
- Researchers call out Palisade AI's LLM self-replication claims as wildly exaggerated — jessi_cata · 2026-10-20
- Wikimedia's agent incident: no permission contract, so agents caused a Wikidata outage — usehive · 2026-10-20
- stealthbench adds 10 EDR-guarded tasks: models' opsec behavior is abysmal — dyn___ · 2026-10-20
- Agentic vulnerability discovery across Windows drivers, XNU, firmware, and web apps — moyix · 2026-10-20