The International Centre for Investigative Reporting (ICIR) is an independent, non-profit newsroom in Nigeria committed to strengthening accountability through investigative journalism.

The problem they wanted to tackle was a persistent challenge in Nigerian journalism: the lack of AI tools that understand and serve local languages.

Their solution, NativeAI for Newsrooms, a culturally aware AI-powered tool that transcribes audio-visual content into text and translates it into Nigeria’s three major local languages: Hausa, Igbo, and Yoruba.

The problem: When foreign AI tools don’t speak your language

For Nigerian journalists, transcription has long been a tedious burden. At ICIR, reporters spent hours manually transcribing interviews, press briefings, and reports - often at the expense of investigative work. Existing tools offered little relief: Whisper, an open-source model developed by OpenAI, required technical fine-tuning before use, while Google Cloud Speech-to-Text was both complex and costly.  

Translation posed its own challenges. Many existing models frequently fell short in accuracy, particularly when it came to capturing the nuances of the Nigerian English accent and inflections, and often produced inaccurate outputs in local languages. 

“Accent differences are a major challenge,” explained Chukwudi Iwuoha, ICIR’s Senior Programmes Officer. “That was why we needed a tool that could adapt to Nigerian languages and accents.”

But the problem extended beyond journalists. Nigeria’s deaf community, supported by growing efforts to promote inclusion and expand access to education, continues to face barriers in accessing transcriptions of audio-visual contents. At the same time, many grassroots media outlets publishing in local languages lack reliable tools to produce accurate translations.

These frustrations sparked a conversation within ICIR about building a solution that was local, simpler to use, and free.

Building the solution: An African-born AI tool

ICIR set out to build something different - an African-led, culturally aware AI transcription and translation tool. They needed to train their AI model to understand and adapt to the unique characteristics of the Nigerian and African accent. With support from the JournalismAI Innovation Challenge, they developed NativeAI for Newsrooms, designed specifically to:

• Transcribe audio-visual content automatically

• Adapt to Nigerian accents and speech patterns

• Accurately translate English transcriptions into Hausa, Igbo, and Yoruba

• Support inclusivity by making content accessible to the deaf community

• Promote the use of local languages in communication, research and teaching

“Our solution serves newsrooms, organisations, and individuals seeking efficient transcription services,” said Iwuoha. “It also supports the deaf community, who may require translation of audio-visuals into readable texts that help bridge gaps to access and bring more people into national conversations.”

From proposal to a working prototype

The journey from idea to prototype was collaborative all the way from the start. The project kicked off with the entire ICIR team aligned on a shared vision. “It was all hands on deck from the very beginning,” recalled Iwuoha. “Everyone contributed - from the proposal stage to the final build. We deliberated extensively to ensure the project matched exactly what we needed.”

The team was cross-functional, bringing together machine learning and data engineers and a full-stack developer, alongside journalists, editors, and fact-checkers who ensured the outputs met editorial standards. Partnerships were essential, as noted by Iwuoha:

“We wanted this to be African-built, for Africa. So we engaged not just our newsroom but external engineers and partners who could provide the expertise and datasets needed.” The ICIR team sketched out a roadmap.

A key milestone was model selection. After testing several options, and with input from external engineers, the team chose Whisper ASR, an open-source speech recognition model, because of its flexibility, speaker differentiation and potential to be fine-tuned for Nigerian English accents and reduce background noise. For translations, they adopted M2M-100, Meta’s multilingual model, which could render accurate outputs in Hausa, Igbo, and Yoruba while handling diacritics and code-switching to preserve meaning.

To ensure reliability, the models were deployed on Google Cloud with HTTPS protocols for secure data handling. Journalists played a central role, uploading interviews then providing feedback that guided improvements in accent recognition and translation quality. The NativeAI model currently achieves approximately 90% accuracy in transcription.

The roadblocks that tested the ICIR team

Like any innovation project, NativeAI faced its share of hurdles - the biggest one being transcription of large files. Initially, the tool struggled to transcribe files longer than 20 minutes, sometimes taking over an hour to process, which defeated the purpose of efficiency. This was resolved after data engineers fixed a backend–frontend communication error. Now, shorter files (5 – 15 minutes) transcribe in just 1 – 2 minutes, while hour-long recordings take 5 – 7 minutes. Translation is even faster, averaging about 3 minutes regardless of file length.

Other key challenges included:

• Data scarcity: Reliable, structured datasets in Nigerian languages were limited, requiring the team to spend considerable time cleaning and preparing data.

• Noisy environments: Street interviews and crowded press events continue to affect transcription accuracy.

• Accent diversity: While the model performs well with Nigerian English, it still struggles with uncommon words and regional name variations.

• Connectivity: Poor internet in some regions disrupts performance, reflecting infrastructure challenges beyond the tool itself.

The opportunity: Expanding NativeAI across Africa

NativeAI may have started as a prototype, but its story doesn’t end there. ICIR plans to refine its accuracy, expand its language coverage, and explore partnerships with other newsrooms and institutions. “This model will not stop in Nigeria. It was made in Africa, by Africa, and for Africa. Expanding to include other widely spoken languages, like isiZulu [a language spoken in South Africa], would make it truly continental and allow other countries to benefit as well” said Iwuoha.

Lessons for newsrooms

From ICIR’s experience, three key takeaways stand out:

• Localisation is everything: AI tools become truly valuable when adapted to local realities - from languages and accents to the specific needs of journalists in their context.

• Collaboration is key: NativeAI came to life because ICIR combined the skills of journalists, data engineers, and external partners. The blend of newsroom insight and technical expertise was essential to solving the problem.

• AI is an aid, not a replacement: Even the best models need human oversight. Journalists remain indispensable in verifying accuracy, providing context, and upholding ethical standards.