In this interview with Impact Newswire, Ashish Dsa, co-founder and CTO of New York-based Voice AI company Arbor, discusses whether recent advances in large language models, speech recognition and generative AI are finally making Voice AI practical for Africa’s diverse linguistic landscape. While global investment in voice assistants and conversational AI has accelerated, most progress has been concentrated in high-resource languages such as English, Mandarin, Spanish and French. Africa, with more than 2,000 languages, limited speech datasets, diverse dialects and widespread code-switching, continues to pose significant technical challenges for Voice AI developers.
Voice AI has entered a new phase following rapid advances in large language models (LLMs), speech recognition and generative AI. Technology companies including OpenAI, Google, Microsoft, Anthropic, Meta and Amazon have significantly improved the ability of AI systems to understand natural speech, generate human-like responses and carry out increasingly complex voice-based tasks. These developments have accelerated investment in voice assistants, AI-powered customer service, healthcare triage, education and workplace productivity tools worldwide.
Yet most of these advances have focused on high-resource languages such as English, Mandarin, Spanish and French, which benefit from vast amounts of text and speech data. Africa presents a markedly different challenge. The continent is home to more than 2,000 languages, many of which have limited digital resources, fragmented dialects and relatively small publicly available speech datasets. As a result, Voice AI systems often struggle with recognition accuracy, pronunciation, accents and multilingual conversations.
Recent advances in multilingual LLMs, self-supervised speech models and synthetic data generation have improved support for low-resource languages, prompting researchers and startups to revisit whether Voice AI can finally scale across Africa. Companies are increasingly experimenting with speech technologies for local languages including Swahili, Yoruba, Amharic, Hausa, Zulu and Wolof, while governments, universities and open-source communities are building larger speech corpora to close longstanding data gaps.
One of the defining characteristics of African communication is code-switching, where speakers naturally alternate between English, French, Arabic, Portuguese and indigenous languages within a single conversation. Developing Voice AI that can seamlessly understand these multilingual interactions remains one of the industry’s most difficult technical challenges, alongside handling regional accents, dialectal variation and noisy real-world environments.
The commercial opportunity is significant. Voice interfaces could expand digital access for hundreds of millions of people who are more comfortable speaking than typing, particularly in sectors such as healthcare, agriculture, financial services, education, government services and customer support. They also offer the potential to improve inclusion for users with limited literacy or limited access to traditional digital interfaces. However, questions remain about model accuracy, bias, infrastructure constraints, affordability and the availability of representative African language data needed for reliable deployment at scale.
Against this backdrop, Ashish Dsa, co-founder and CTO of New York-based Voice AI company Arbor, shares his practitioner perspective with Faustine Ngila, on how far the technology has progressed. Arbor develops Voice AI systems for frontline workforces, giving Dsa firsthand experience building conversational AI for users who operate outside traditional desk-based environments. Formerly an engineer at Meta and Telus, he has also served as a judge for the MIT $100K Entrepreneurship Competition, reviews AI books for Manning Publications, and has received a Forbes Editor’s Choice Award.
He discusses whether Voice AI is finally becoming practical for Africa’s diverse linguistic landscape, the technical barriers that remain, and what breakthroughs are still needed for widespread adoption. Here is the interview:
1. Large language models have made significant advances in English and other major languages. How well do today’s Voice AI systems perform in African languages such as Swahili, Yoruba, Amharic, Zulu and Wolof?
English still sets the bar for production Voice AI. Of the languages you named, Swahili usually has more speech and text in public research and commercial stacks than Yoruba, Amharic, Zulu, or Wolof, so system quality tracks that data gap. Error rates still jump when the room is noisy, the dialect shifts, or the domain vocabulary is thick. I ship production Voice AI for frontline workers in enterprise settings. Low-resource African languages show up in demos far more often than they hold under real load.
2. What are the biggest technical challenges in building Voice AI for African languages? Is the main bottleneck the lack of training data, speech datasets, dialect diversity, accents, or something else?
Labeled speech data is still the choke point. You need multi-speaker audio in noisy rooms, with domain words and real dialect coverage. Clean studio hours alone will not get you there. Text models can look fine while the speech front end fails on the same utterance. Dialect and accent variation sit next to data volume as first-order problems, and code-mixing makes labeling harder on top of that. After that, eval harnesses have to break scores down by language and by community. One global average hides who the system fails.
3. Many African countries are multilingual. How effectively can Voice AI switch between languages or understand code-switching, where speakers mix English, French or Arabic with local languages?
Still weak in production. Models trained on clean monolingual audio break when a speaker flips mid-sentence between English and a local language, or between French or Arabic and a local language. Some stacks detect language, then route to a single-language model. True code-switch breaks that path. Multilingual end-to-end models help when mixed training data exists, which for many African language pairs is still thin. I design for code-switch as a failure mode from day one.
4. In your view, which sectors stand to benefit the most from accurate Voice AI in local languages?
Anywhere the user is not at a desk and typing is a poor fit: healthcare follow-up, agriculture extension, mobile money and banking support, government citizen services, voice-first customer service. I build Voice AI that interviews frontline workers at scale, so I lean toward speech as the natural channel. Local-language accuracy decides whether the product reaches the full population or only the English-comfortable slice.
5. How important is Voice AI in improving digital inclusion for people with limited literacy or those who are more comfortable speaking than typing?
High. For people with limited literacy, or who prefer speaking to typing, voice is the path in. On our product, frontline workers complete voice interviews at rates above 95 percent, well above what we see from legacy text and paper surveys. That pattern is about modality and being heard. The same logic applies where keyboards and formal literacy get in the way.
6. Can you share any examples or case studies where Voice AI in an African language has delivered measurable improvements in productivity, customer experience or service delivery?
I will not invent one. Arbor runs production Voice AI interviews with frontline workers at scale. We do not have an African-language deployment I can put on the record as a case study. Happy to talk through production failure modes that transfer, or to review a public study you want carefully attributed. I will not cite outcome numbers I have not audited.
7. How do you ensure Voice AI systems accurately understand regional accents and dialects without introducing bias or excluding certain communities?
Continuous measurement. Build eval sets sliced by region, dialect, and community, then track error by slice rather than only the headline average. Collect more data for under-covered accents instead of averaging them out of the score. Human review and LLM-as-judge eval harnesses help when labels cost a lot, but the ground truth still has to include the people you serve. If a dialect never appears in training or eval, the model fails it quietly.
8. How close are we to Voice AI that can hold natural, reliable conversations across multiple African languages, and what breakthroughs are still needed to get there?
Closer for narrow domains and better-resourced language pairs. Not close for natural, reliable multi-turn conversation across many African languages in noisy field conditions. Still missing: large licensed speech corpora with dialect labels, stronger code-switch models, paths that work on low bandwidth, and eval culture that refuses an English-only score. Production agent harnesses and multi-agent orchestration help once speech is solid. A weak speech front end stays weak underneath them.
Faustine Ngila is the AI Editor at Impact Newswire, based in Nairobi, Kenya. He is an award-winning journalist specializing in artificial intelligence, blockchain, and emerging technologies.
He previously worked as a global technology reporter at Quartz in New York and Digital Frontier in London, where he covered innovation, startups, and the global digital economy.
With years of experience reporting on cutting-edge technologies, Faustine focuses on AI developments, industry trends, and the impact of technology on society.
Danai Gurira (Zimbabwe/American)
Most Famous Movie: Black Panther (2018)
Short Bio: Born in Iowa, USA, but raised in Zimbabwe, Danai Gurira is an acclaimed actress and playwright. She is widely known for her powerful portrayal of Michonne in The Walking Dead and as Okoye in the Marvel Cinematic Universe films, particularly Black Panther.
All Natural
El Salvador: You have two cows. They are less likely to get cut open than you.
Necessary Momentum
Finland: You have two cows. They are more likely than your entire country to commit a crime.