AI researchers now want search tools that recognize not just what was said, but who said it.
A paper posted to arXiv on September 17, 2026 details VoiceTrace, a two-stage retrieval framework, and VoiceTrace-Bench, a companion benchmark for what its authors call hybrid speech retrieval. Each query in the benchmark combines text describing what to find with a reference audio clip specifying whose voice to find it in. VoiceTrace works in two steps: VoiceTrace-Emb builds unified embeddings for fast retrieval across large audio collections, then VoiceTrace-Reranker examines each query-candidate pair more closely to sharpen the results. The researchers report that VoiceTrace also achieves state-of-the-art results on existing, speaker-agnostic semantic speech retrieval benchmarks.
That's the useful part. Searching meeting recordings, podcasts, or lecture archives for a topic is a solved problem. Searching them for a topic raised by one specific person is not, and most existing tools treat speaker identity as an afterthought if they handle it at all.
Whether VoiceTrace-Bench catches on as a standard, rather than a one-paper benchmark cited mostly by its own authors, is the part the abstract can't answer.