SPARC tests speaker linkage from speech transcripts alone
Original title: SPARC: Spoken Profile Authorship Recognition Corpus
LLMs separate same-speaker pairs from text; 50-word probes weaken the strongest adversaries most.
SPARC studies how much speaker information remains in transcribed words after acoustic anonymization. The corpus is built from two podcast sets with more than 12,500 speakers and 1.5 million quality-filtered utterances, plus sex, age, and accent labels. Five attribution models and three LLMs are evaluated on identity and attribute trials. LLMs distinguish same-speaker pairs from text, while sex and accent verification exceed chance.
- The corpus covers over 12,500 speakers and 1.5 million utterances.
- Labels cover sex, age, and accent, inferred or source-provided.
- Attribute-label agreement is harder to verify than speaker identity.
- Releasing only 50-word probes degrades the strongest adversaries most.
From the original article
1 min read at huggingface.co
Speech anonymization has focused on the acoustic signal, leaving open how much speaker information the transcribed words themselves retain. In this work, we study pairwise speaker and demographic-attribute verification from transcribed speech. We introduce SPARC (Spoken Profile Authorship Recognition Corpus), a benchmark built from two podcast corpora comprising over 12,500 speakers, 1.
By Efthymios Tsaprazlis, Bowen Yi, Tiantian Feng · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.