The idea
Papiamentu is spoken by a few hundred thousand people, and off-the-shelf speech AI barely understands it. So before promising anything, we measured.
Context
An organisation in Curaçao holds its meetings in Papiamentu. Minutes are written by hand, depend on one note-taker's reading of the discussion, and arrive late. Decisions lose their reasoning, and action items lose their owners.
The challenge
Most speech AI is trained on large languages. Papiamentu has very little training data, so the real question was not which tool to buy but whether reliable transcription is possible at all, and what it would take to get there.
Our strategy
- 01Built and deployed a working Papiamentu speech-to-text prototype and tested leading models on real native-speaker audio
- 02Found that a popular default model got 95% of words wrong, while the best candidate got only 14% of characters wrong: it hears the sounds and mostly misspells them
- 03Traced most remaining errors to word-joining and spelling conventions, which fine-tuning on a few hours of consented meeting audio can fix
- 04Designed the platform around that finding: one microphone per participant, a transcript the note-taker reviews, then summaries, decisions and action items with owners, each linked back to the transcript
- 05Planned the delivery as twelve months of two-week sprints, with a proof-of-concept gate that fixes the accuracy commitment on the client's own audio before the build continues
What we delivered
- Working Papiamentu speech-to-text prototype
- Benchmark of speech models on real native-speaker audio
- Platform architecture: per-speaker capture, reviewed transcript, structured meeting record
- Twelve-month delivery plan with a proof-of-concept accuracy gate
- Team plan: 3 core members and 4 specialists, including a Papiamentu language lead
Where it stands
Instead of a promise, the plan starts from measured evidence: what today's models get right in Papiamentu, what they get wrong, and the shortest path to a transcript people can trust.