New AI Search Finds Who Said What in Any Recording

A team of researchers has published VoiceTrace, a new way for computers to search through recorded speech. Today, if you want to find a moment in a meeting, podcast, or interview, you either scrub through the audio by hand or search a written transcript for keywords. VoiceTrace adds a second filter: who was speaking. You hand it a short clip of a person's voice, type what you're looking for, and it returns the moments that match both.

The clever part is that you don't need to type a name. In real life, you often recognize a voice without knowing who it belongs to — think of a panel discussion, a customer support call, or an old family recording. The system takes your reference clip and compares it against every voice in the archive. It then ranks results in two stages: a fast first pass that narrows thousands of hours down to a short list, and a slower second pass that carefully re-checks each candidate before ranking it.

The team also built VoiceTrace-Bench, a test set for this exact task, because older speech-search tests only checked whether a system found the right words and ignored whether it found the right person. On their new test, VoiceTrace beat older approaches that handled the two jobs separately. That matters because 'who said it' is often the whole point — a quote is worth very little if you can't attribute it to anyone.

Here's the catch. This is a research paper, not an app you can download today. And a voice is a biometric — as personal as a fingerprint. A tool that can identify speakers across recordings could also be used to track people without their consent, so the same technology that helps you find a quote could help someone else find you. Expect the useful and the uncomfortable uses to arrive together.