SPARC: Spoken Profile Authorship Recognition Corpus

SPARC. Spoken Profile Authorship Recognition Corpus. By Efthymios Tsaprazlis and five others, from the University of Southern California and the University of Michigan. Published on the eighteenth of October. Its code repository is public and real: about twenty source files.

Work on anonymizing speech has concentrated on the acoustic signal, the voice itself. That leaves an open question. Once speech is transcribed, how much about the speaker do the words alone still give away?

The authors study that as pairwise verification from transcribed speech. Given two pieces of text, were they spoken by the same person? And do the two speakers share a demographic attribute, such as sex, age or accent?

SPARC, the Spoken Profile Authorship Recognition Corpus, is the benchmark built for this. It draws on two podcast corpora, and covers more than twelve thousand five hundred speakers and one point five million utterances that passed quality filtering. Each speaker carries labels for sex, age and accent, either inferred or provided by the source. Five attribution models and three large language models are evaluated on trials that verify identity and attributes, with topic proxies based on episodes and on metadata.

Large language model adversaries can tell same-speaker pairs apart from text alone. Agreement on attribute labels is harder to verify, but the language models still verify sex and accent labels above chance. Finally, releasing only short probe texts of fifty words degrades the strongest adversaries the most, which gives a partial reduction in observed linkage.

Stripping the voice, then, is not the same as removing the speaker. Text alone links people back together, and shortening the released text helps only in part.