Speech anonymization systems typically alter the acoustic properties of an audio recording to hide who is speaking. But according to a new arXiv paper, this leaves an open question: how much speaker information do the transcribed words themselves retain? The paper introduces SPARC, the Spoken Profile Authorship Recognition Corpus, to study exactly that.
The authors propose pairwise speaker and demographic-attribute verification as a way to measure what can be inferred from transcripts alone. This shifts the research lens from the sound of a voice to the content and style of what was said. The corpus is designed to support such verification tasks, offering a benchmark for a dimension of privacy that has been largely overlooked.
Because this is a single source, there are no conflicting findings to compare. The paper's contribution is the corpus itself and the framing of a new research direction: treating transcribed speech as a potential carrier of identity and demographic information. The abstract does not report specific results, so the practical extent of the leakage remains an open empirical question.