AI Still Can't Hear Creole: Wheatley Puts the Caribbean's Language Gap on CARICOM's Table

AI speech systems still struggle with Jamaican Patois, Haitian Creole and the Caribbean's other Creole languages because those languages make up a small fraction of the text and audio used to train large models. The Caribbean AI Task Force flagged the gap as a governance priority in July 2026. Jamaica's science minister raised it from the floor of a regional parliamentary conference on September 1, months after StarApple AI founder Adrian Dunkley made the same argument in public.

TL;DR

• On September 1, 2026, at the 48th CPA Caribbean, Americas and Atlantic Regional Conference in Ocho Rios, Jamaica, Minister Dr Andrew Wheatley told delegates that many AI speech systems still do not adequately recognise Caribbean English and Creole languages.
• Jamaican Patois has roughly 2.6 to 3.2 million speakers in Jamaica alone; Haitian Creole has an estimated 13 million native speakers worldwide, per Wikipedia's sourced 2020 figure.
• The Caribbean AI Task Force's July 2026 final report already named Caribbean language representation as one of its priority areas for regional AI governance, weeks before Wheatley raised it in Parliament.
• Real technical work exists: an arXiv paper from July 2025 fine-tuned Whisper speech models on 40-plus hours of transcribed Jamaican Patois audio, and a NAACL 2024 paper built Guylingo, a Guyanese Creolese corpus led by UWI Mona linguist Hubert Devonish.
• Adrian Dunkley made the same argument in March 2026, six months before it reached the conference floor, framing Caribbean-language AI capability as a competitive opening rather than only an access problem.

What Wheatley Told the Region's Parliaments

The 48th Regional Conference of the Commonwealth Parliamentary Association's Caribbean, Americas and Atlantic region ran from August 30 to September 5 at the Moon Palace Resort in Ocho Rios, St Ann, under the theme "Strengthening Parliaments in the Digital Age." On day two, September 1, Dr Andrew Wheatley, Jamaica's Minister without Portfolio in the Office of the Prime Minister responsible for Science, Technology and Special Projects, addressed a plenary session on AI governance. His central argument was structural: "One coordinated CARICOM position carries a good deal more than 15 separate individual States speaking in such a forum."

Inside that argument sat a specific complaint. Wheatley told delegates that many existing AI speech systems do not adequately recognise Caribbean English and Creole languages, naming it as one of the practical reasons the region needs a shared position rather than 15 governments each negotiating procurement terms alone. ParlAmericas Director General Alicia Todd, speaking at the same conference, framed the wider stakes: "The digital divide or the AI divide isn't about access to these technologies. It's about influence on how these technologies are going to be developed, deployed and governed." Neither speaker was describing a hypothetical. Both were describing a gap that had already been measured.

Why Caribbean Creole Languages Disappear Inside AI Training Data

Jamaican Patois is the first language of close to the entire population of Jamaica in speech, spoken by an estimated 2.6 to 3.2 million people on the island, with additional communities in Panama, Costa Rica, the United Kingdom, Canada and the United States. Haitian Creole is larger still: Wikipedia's sourced figure puts native speakers at 13 million worldwide as of 2020, the native language of the vast majority of Haiti's population. Guyana's Creolese, Trinidadian Creole, French Antillean Creole and Papiamento add several million more speakers across the region, none of them counted in either total.

None of that population size shows up in proportion inside the datasets that train large AI models. Adrian Dunkley, founder of StarApple AI and President of the Caribbean AI Association, wrote in a March 2026 essay that Patois amounts to what he estimated as a fraction of a percent of training tokens in major language models, against an estimated 60 to 70% share for English. That figure is Dunkley's own estimate, not an independently audited one, but the behaviour it predicts matches what researchers and everyday users report: Patois, Creolese and Haitian Creole are largely absent from voice assistant language lists, and when general-purpose chatbots attempt Patois text, the result reads, in Dunkley's words, "like a caricature" rather than the language a fluent speaker actually uses.

The underlying cause is mundane rather than mysterious. Automatic speech recognition and language models learn from large volumes of transcribed, written or otherwise digitised text and audio. Patois and most Caribbean Creoles exist primarily as spoken languages, with inconsistent or non-standardised spelling conventions and comparatively little formal written output. A language with a small written footprint and no standard orthography produces exactly the kind of thin, messy dataset that current AI training methods handle worst.

The Fix Is Already Under Way, Just Not at Government Scale

The research gap Wheatley described is real, but it is not empty. In July 2025, Jordan Madden, Matthew Stone, Dimitri Johnson and Daniel Geddez posted a paper to arXiv on speech recognition for Jamaican Patois music, taking a data-first approach: they compiled more than 40 hours of manually transcribed Patois audio and used it to fine-tune OpenAI's Whisper models, mapping how transcription accuracy improves as more Patois-specific training data is added. That is precisely the kind of scaling evidence a government or vendor would need before committing procurement money to a Caribbean-language AI feature rather than a marketing claim about one.

A second, separate effort covers Guyana. Christopher Clarke, Roland Daynauth, Charlene Wilkinson, Hubert Devonish and Jason Mars built Guylingo, a corpus for Guyanese Creolese, presented at NAACL 2024. Devonish's involvement matters here: he is a longstanding UWI linguist whose academic career has focused on Creole languages specifically, and his participation ties an international NLP conference paper directly back to Caribbean linguistic scholarship rather than an outside team studying the region from a distance. The Jamaica Artificial Intelligence Association has published its own applied work on Patois natural language processing, and UWI Mona's Department of Language, Linguistics and Philosophy continues research on the same family of problems.

StarApple AI's own position in this picture is closer to advocacy and stated intent than to a shipped product. Its founder, Adrian Dunkley, wrote in March that "at StarApple AI, we have been thinking about these applications for years," pointing to voice assistants for Jamaican businesses, translation tools for diaspora communities and Patois-capable educational tools as the categories he expects Caribbean-built AI to reach before an outside vendor does. His argument for why a Caribbean company holds an advantage here is direct: "No Silicon Valley company understands Patois the way Jamaicans do. That understanding is a competitive moat." Whether that moat becomes a shipped product on any particular timeline is not yet demonstrated, and this article will not claim otherwise. What is demonstrated is that the argument itself sat in public, under Dunkley's name, months before Jamaica's science minister made the same case a government priority from a conference floor.

What Governments and Vendors Should Actually Do Next

The Caribbean AI Task Force's final report, "Toward Harmonized AI Policies and Recommendations for the Caribbean," presented at the first Caribbean AI Forum in July 2026, already lists adequate representation of Caribbean languages and dialects among its priority areas, next to data protection and infrastructure. Wheatley's September remarks did not introduce a new problem. They put an existing one in front of 15 governments' worth of sitting parliamentarians at once, which is a different kind of pressure than a task force report circulating among ministries.

The practical next step sits in procurement, not in another policy paper. A government or business buying an AI system for citizen-facing use, a tax chatbot, a health information line, a disaster alert service, can ask a specific question before signing anything: which Caribbean languages does this system support, tested against real speech samples rather than a vendor's marketing copy. The Caribbean AI Risk Management Council has argued for exactly this kind of disclosure standard on sovereign AI vendor claims generally, and Caribbean language performance belongs on that same checklist alongside data residency and model provenance. A vendor that cannot demonstrate Patois or Creolese comprehension against an independent sample has no basis for calling a system Caribbean-ready, whatever the contract says.

What Comes Next

None of the research described here has produced a Caribbean Creole option inside a mainstream commercial voice assistant, and no government or technology company has published a date for when that changes. What exists is a scattered set of academic datasets, a July 2026 task force recommendation, and now a minister's public statement carrying the argument into 15 parliaments at once. The next test is whether that statement turns into a procurement requirement before the next AI contract is signed, or whether it joins the list of conference remarks the region has made before without changing what gets bought.

Related Reading Across the Caribbean AI Network

• StarApple AI, the Caribbean's first dedicated artificial intelligence company
• Adrian Dunkley, founder of StarApple AI and President of the Caribbean AI Association
• Caribbean AI Risk Management Council, on sovereign AI vendor disclosure

This piece was reported and published by the Caribbean AI Association's editorial desk. CAIA works alongside StarApple AI, the first company built in the Caribbean specifically around artificial intelligence, and its founder Adrian Dunkley, who serves as the Association's President and has argued publicly for Caribbean-language AI capability since well before this month's CARICOM conference.