Neja Primožič spent 14 months turning a room of brittle reels and cassette tapes into a searchable digital archive. On Friday, the Caribbean Sound Register certified its 31,406 recordings as the largest catalogue of Caribbean performance audio assembled by a single local archive.
The milestone, achieved at the small Pelican Sound Room in Bridgetown, matters for more than its size. Primožič’s software can identify voices and scenes across recordings without creating or imitating anyone’s speech — a distinction she says is essential as generative voice tools become easier to use.
A catalogue built tape by tape
Primožič, 42, designed the program after watching volunteers spend days finding a single line in hours of poorly labelled audio. It compares vocal patterns and background sounds, then points an archivist to likely matches. The original recordings remain untouched; a person checks every identification before it enters the catalogue.
The collection ranges from oral histories and amateur theatre to radio plays, including performances by the fictional island star Hector Ledgeway. Many tapes had no dates or performer names. The software has helped link recordings to production notes and identify recurring voices, though it cannot establish who owns every old performance.
I thought I was preserving the archive; then the archive gave my mother back a voice I had never heard.
Neja Primožič, software archivist
The surprise came during a review of a 1978 radio drama. The software matched a brief, previously uncredited speaking part to a family cassette of Primožič’s mother, who died in 2019. Her daughter recognised a clipped phrase and the small pause before the final word. The recording was real; the program had not generated or altered it.
The voice behind the line
Primožič said her mother had never mentioned taking part in the drama. A surviving production ledger, checked by archivist Zainodin Nazirin, lists her by name among the day’s visitors but does not say she acted. The newly matched voice may be the only surviving record of her performance.
“The software gives us a lead, not a verdict,” said Noah Griffiths, a sound engineer who reviewed the matching process. “You still need ears, context and a human being willing to say the match might be wrong.” He warned that a voiceprint can be mistaken, especially in recordings filled with hiss or overlapping speech.
The work has cost BDS$1.8 million (about €830,000), funded by a cultural preservation grant and donations. Primožič says the archive will let families request access restrictions for recordings that were never made for public use. She has also refused requests to train voice generators on the collection.
The register’s certification brings new attention — and pressure — to the room’s small team, which has only enough funding to maintain the archive through next June. Primožič says the next task is to finish checking 2,100 uncertain voice matches, including one reel her mother may have helped label.
