Latest

We use a well established process to develop Speech Recognition technologies that transform the future of communication.

Man in patterned shirt sitting in a classroom gesturing with one hand while holding an assistive speech device to his chin.
Some voices are left out of the algorithm. We’re working to change that.
Dr. Richard Cave

Our Process

1
on non-standard speech
A woman in a red shirt showing a tablet to a man in a pink checkered shirt inside a room with chairs and curtains.

We begin with the collection of 'non-standard' speech data with support from people living with impaired speech - talking in the languages they speak - in each country we work in. These recordings are used to train AI-driven Automated Speech Recognition (ASR) models that can better recognise impaired speech. Then, we train people on how to use applications powered by these models, and we continue to fine train our models through real world usage and feedback.

Download our data
2
to recognise that speech

Our ASR models are built using self-supervised learning and transfer learning techniques, allowing us to develop effective speech recognition even with limited initial data. We leverage pre-trained multilingual models and adapt them specifically for underrepresented languages and non-standard speech patterns. By training on the rich, diverse datasets we've collected, our models learn to recognise unique linguistic features, tonal variations, and speech characteristics that make each language distinct, creating AI that genuinely understands how people actually speak, regardless of where they live.

Our Research
3
to use apps powered by these models
Classroom scene with a teacher standing and explaining to attentive students seated at tiered desks with laptops and notebooks.

Technology is only transformative when people know how to use it. We build communities of innovators, users, and ecosystem enablers through hands-on training programmes. Working with speech therapists, educators, developers, and users themselves, we provide training on how to use ASR applications, contribute to data collection, and understand the technology's potential. This ensures that the solutions we create don't just exist—they're actively used, maintained, and championed by the communities they serve, creating lasting digital inclusion wherever communication barriers exist.

Our Apps
4
we run local hackathon events
Group photo of participants and organizers standing on a red carpet at Tekyerema Pa Hackathon event with a woman speaking at a podium on the left.
This hackathon is not the end, but it's rather the beginning of what comes next.
Prof. Isaac Wiafe

The final piece of the puzzle is local innovation sprints. These extended collaborative events bring together developers, linguists, speech therapists, and users alongside community partners, academic institutions, and local technology ecosystems to develop applications using our open-source datasets and models. Through structured support and deep engagement, Innovation Sprints generate practical solutions tailored to local needs, from communication apps to educational tools, while simultaneously building lasting technical capacity within communities.

Building on successful sprints in Ghana and Kenya, with Rwanda planned for the near future, we're expanding this model globally. Beyond rapid prototyping, these Innovation Sprints create pathways for local ownership of AI solutions, strengthen regional technical expertise, and provide evidence of what's possible when communities have the tools and support to build inclusive AI applications that directly address their linguistic and accessibility needs.

Discover our latest news and events

2025: Africa

Our work began in Africa, where millions of people with non-standard speech have had little access to assistive technology or speech therapy—and where most languages have been entirely absent from AI development. Starting with a successful pilot in Ghana, we've expanded across the continent, conducting innovation sprints in Kenya, Rwanda, and beyond.

Read more from our work in Africa
Close-up of a person's face, showing their contemplative eyes and a large gold earring.
Radia's Story
Close-up profile of a man with short hair and a shaped hairline, looking contemplatively to the right.
Oscar’s Story
Map of Africa, showing Ghana, Uganda and Kenya highlighted.

In each location, we partner with local speech therapists, linguists, and communities to collect authentic speech data in native languages. These aren't just vocabulary lists—they're the words and phrases people actually use in their daily lives. Through intensive hackathons, we bring together local developers, researchers, and users to rapidly prototype applications that address real communication challenges. The result is a growing network of innovators who understand both the technology and the communities they serve, creating solutions that are culturally relevant, linguistically accurate, and built to last.

View progress on our interactive map

Outcomes

​

Datasets & models

All our data is open source and freely available. We provide comprehensive datasets of non-standard speech in multiple languages, alongside trained ASR models to better recognise diverse speech patterns. These resources enable researchers, developers, and innovators worldwide to build inclusive communication technologies, systems and services.

Explore our datasets and models
​

Apps

We develop and support open-source applications to put our multi-language speech recognition technology into practice. Alongside local developers and communities, we co-create mobile apps and tools that enable rapid adoption of open-source non-standard speech data to facilitate AI-powered speech recognition.

Learn more about our apps
© 2026 CDLI. All rights reserved. Site by Fry
Accessibility
Accessibility, A11y
The design of technologies, services, and environments so they can be used by people with diverse abilities, including people with disabilities.
Assistive Technology
Assistive Technology
Tools or systems that help people with disabilities perform tasks more easily. Inclusive speech recognition can function as assistive technology for communication.
Automatic Speech Recognition (ASR)
Automatic Speech Recognition (ASR), Automatic Speech Recognition, ASR
Technology that converts spoken language into written text. ASR enables devices and software to "understand" human speech and is used in tools such as voice assistants, transcription services, and accessibility technologies.
Community-Centred Design
Community-Centred Design, Community Centred Design
An approach that actively involves people who will use the technology in its design, development, and testing, ensuring solutions reflect real needs and lived experiences.
Data Annotation
Data Annotation
The process of labelling speech recordings with accurate text or metadata. Annotation helps machines learn how spoken words relate to written language.
Digital Language Inclusion
Digital Language Inclusion
The practice of ensuring that digital technologies recognise, understand, and support all ways people speak, including diverse languages, accents, and speech patterns.
Ethical Data Collection
Ethical Data Collection
The practice of collecting data with informed consent, transparency, privacy protection, and respect for participants' rights and dignity.
Inclusive AI
Inclusive AI
Artificial Intelligence systems designed to work fairly and effectively for people of different abilities, languages, and backgrounds, without excluding or disadvantaging certain groups.
Machine Learning
Machine Learning
A branch of AI that allows computer systems to learn patterns from data and improve their performance over time without being explicitly programmed for every task.
Non-Standard Speech
Non-Standard Speech
Speech that differs from what most speech recognition systems are trained on. This may include speech affected by disability, neurological conditions, injury, accent variation, or local language patterns.
Open-Source
Open-Source, opensource, open source
Software, data, or models that are freely available for anyone to use, study, modify, and share. CDLI promotes open-source approaches to accelerate inclusive innovation.
Speech Dataset
Speech Dataset
A structured collection of recorded speech samples and corresponding text. These datasets are used to train and improve speech recognition models.
Speech Impairment
Speech Impairment
A condition that affects a person's ability to produce clear or typical speech. This can be temporary or lifelong and may result from disability, illness, injury, or neurological differences.
Speech Model
Speech Model
A trained AI system that processes and recognises spoken language. Speech models improve as they are exposed to more diverse and representative speech data.
Under-Resourced Languages
Under-Resourced Languages, Under Resourced Languages
Languages that lack sufficient digital data, tools, or technological support. Many African languages fall into this category, making them poorly supported by mainstream speech technologies.