Krio AI

What is Krio?

Krio is an English-lexifier creole language spoken in Sierra Leone. It is the first language of the Krio community — descendants of freed and repatriated Africans who settled around Freetown from the late 18th century — and, more importantly for technology, it is the common language used across ethnic groups throughout the country. Its vocabulary is largely English-derived, layered with Yoruba, Portuguese, other African languages and indigenous Sierra Leonean languages, and it has its own phonemic orthography. Its ISO 639-3 code is kri.

Krio belongs to the same family of West African English-lexifier creoles as Nigerian Pidgin, Ghanaian Pidgin, Cameroonian Pidgin and Aku in The Gambia, so work on Krio has a natural path to those languages.

Why Krio is underserved by AI

Modern speech and language models are trained on very large amounts of digitised text, transcribed audio and parallel translations. Krio has little of any of these. It is widely spoken but under-represented online, has few annotated corpora, no widely adopted evaluation benchmark, and is frequently missing from — or poorly covered by — multilingual systems such as large speech models and translation models.

Several properties make it harder still: pronunciation and spelling vary by region and speaker, everyday Krio mixes freely with English (code-switching), and most real audio is conversational — voice notes, phone calls, radio — rather than clean read speech. Systems tuned on high-resource, studio-quality data degrade sharply under these conditions. Building for Krio means building the data and the evaluation, not just fine-tuning a model.

Capabilities

Krio datasets

Every model starts with data. NovaxAILab collects Krio speech and text from native speakers and public sources, then cleans, annotates, aligns and quality-checks it through a documented pipeline. See the data pipeline for how corpora are built, and research for the wider programme.

Krio AI models

NovaxAILab is building a family of Krio models: automatic speech recognition (in development), a released Krio text-to-speech model on Hugging Face, and Phos, which combines transcription and translation in one workflow. Browse them on the models page. Evaluation results are being prepared for publication on the individual model pages.

Applications

• Accessibility tools that let people use voice in the language they actually speak
• Customer support and call-centre transcription for Krio-speaking users
• Education and literacy tools that read and speak Krio
• Public-service and health information delivered as spoken Krio
• Media and broadcast captioning for Krio-language content

Frequently asked questions

What is Krio AI?

Krio AI refers to speech and language technology built for Sierra Leonean Krio: automatic speech recognition (speech-to-text), text-to-speech, machine translation, and the datasets, tokenisers and language models these systems depend on. Because Krio has little published training data, most of this technology has to be built specifically for the language rather than inherited from large multilingual models.

Does ChatGPT understand Krio?

Partly. General-purpose assistants can often read and produce simple Krio because much of its vocabulary derives from English, but quality drops on idiomatic, spoken or code-switched Krio, and Krio is not listed among the officially supported languages of the major commercial assistants. Speech features (transcribing or speaking Krio) are largely unsupported. This is the gap NovaxAILab works on.

What is Krio speech recognition?

Krio speech recognition, or Krio ASR, is the task of automatically converting spoken Krio audio — voice notes, phone calls, broadcasts — into written Krio text. It is difficult for off-the-shelf systems because they were trained with little or no Krio audio and do not handle Krio phonology, orthography or English–Krio code-switching well.

Is there a Krio speech dataset?

Public Krio speech data is scarce. NovaxAILab collects and curates Krio speech and text through a documented data pipeline covering collection from native speakers, cleaning, annotation and quality control. Dataset release details are published on the datasets and data-pipeline pages as corpora become available.

Can AI translate English to Krio?

Yes, to a degree. Rule-of-thumb translation is possible because of shared vocabulary, but faithful translation that preserves tone and meaning needs parallel English–Krio text and evaluation against Krio references. NovaxAILab's translation research focuses on building that parallel data and measuring quality rather than assuming it.

Can AI generate Krio speech?

Yes. Neural text-to-speech can produce natural-sounding spoken Krio when trained on clean Krio recordings with accurate transcripts. NovaxAILab has released a Krio text-to-speech model and continues to expand the underlying speech corpus.

What is a low-resource language?

A low-resource language is one without the large volumes of digitised text, transcribed audio and parallel translations that modern AI models are trained on. Krio is low-resource: it is widely spoken but under-represented online, has limited annotated corpora, and is often missing or poorly covered in multilingual speech and translation systems.