Africa rarely speaks one language at a time. A doctor might start explaining a diagnosis in English and switch mid-sentence to Swahili or Yoruba to reassure a patient. A bank customer begins in one local language and finishes in English. Friends, families, markets, courts, and clinics blend languages fluidly every day. Yet most voice AI systems still treat this natural code-switching as noise or failure.
That gap is exactly what the Sahara CodeSwitch Africa Challenge 2026 aims to close.
Organised by Intron (the African voice AI company behind the Sahara models) in collaboration with the MLC (Africa) Agentic AI Workshop, the challenge invites developers, researchers, startups, product teams, students, NGOs, and builders across Africa and the diaspora to create real voice applications that understand how Africans actually speak.
The total cash prize pool is US$10,000, with additional non-cash benefits including partner visibility, potential pilot opportunities, Sahara API credits, and recognition.
Why This Challenge Matters
Global speech models often struggle with African accents, local languages, background noise, and especially code-switching—the seamless mixing of languages within a single conversation or even sentence. Intron’s Sahara platform was built specifically for African speech, and the new Sahara v2.5 release (in research preview for participants) focuses on bilingual code-switching across 12 African language pairs, with broader language support expanding.
Participants get early beta access to Sahara v2.5 APIs (speech-to-text and text-to-speech), plus open datasets such as AfriSwitch (natural in-the-wild code-switched speech across 14 language pairs) and AfriSwitchCare (a medical code-switching dataset). There is also guidance on benchmarking and eligibility for related NVIDIA masterclass opportunities.
The goal is bigger than any single model: surface honest performance data on code-switched audio, build useful agentic applications, and push the ecosystem toward voice AI that works in real African environments.
The Challenge Task
This is not a pure transcription contest. Submissions must:
1. Use voice to achieve a downstream task (the agentic component) for example, a voice assistant that completes a booking, a triage tool, a form-filler, or any system that takes action on the user’s behalf—not just convert speech to text.
2. Benchmark performance on code-switched audio across at least three speech models, including an Intron Sahara API and at least two others of the team’s choice (global commercial, open-source, local, pretrained, or fine-tuned).
3. Target a specific vertical or use case.
4. Where possible, submit the code-switched audio samples used for testing (with basic metadata: language pair, domain, accent/country, device, noise conditions).
Competition Categories
Five tracks focused on real-world impact:
- Health: Clinical documentation, patient intake, community/maternal health, hospital workflows, frontline care.
- Legal & Public Services: Court transcription, legal aid, citizen services, civic reporting, justice access.
- Fintech, Telco & Customer Experience: Voice banking, KYC, payments, collections, call centres, agent assist.
- Agriculture & Education: Farmer advisory, extension services, tutoring, classroom learning, local-language skills training.
- Other High-Impact Use Cases: Humanitarian response, accessibility, media, commerce, transport, culture, or any meaningful application of code-switching voice AI.
Bonus consideration goes to solutions addressing real problems for meaningful population sizes. Submissions can range from fresh ideas to near-production or production apps clearly designed for real-world use.
Who Can Enter
Open to builders across Africa and the diaspora. Apply solo or as a team (recommended size 2–5 people). Especially encouraged: developers and engineers, AI/ML researchers, startups and product teams, student teams, open-source builders, NGOs and civic-tech teams, and companies with voice or customer-facing workflows. Multidisciplinary teams (including domain experts, linguists, clinicians, designers, etc.) are welcome.
What to Submit
- Short description of the problem and solution.
- Working prototype (short video demo).
- Code or technical documentation.
- Benchmark results comparing Sahara against at least two other models.
- Note on responsible AI (privacy, consent, safety, data use).
Judging Criteria
Applications will be judged based on Five dimensions:
1. Real-world Impact — Meaningful user need for a sizable population.
2. Code-switching Performance — Handling of multilingual and natural code-switched scenarios.
3. Product Quality — Usability, workflow fit, and user trust.
4. Technical Execution — Reliability, integration, latency, robustness.
5. Ethics & Safety — Consent, inclusion, privacy, bias, and respect for user dignity.
Timeline (Key Dates)
- Phase 1 (Indaba sprint): Already underway around early August 2026 at Deep Learning Indaba.
- Main competition build period: Mid-August to 15 September 2026.
- Registration / Expression of Interest: Opened 5 August; closes around mid-September (final submissions due 15 September 2026).
- Finalist selection, Partner Demo Day: Late September.
- Winners announced: 1 October 2026.
Prizes and Benefits
- US$10,000 total cash prize pool distributed across top-performing solutions in the categories.
- Workshop/sprint prizes (smaller amounts for the Indaba phase).
- Access to Partner Demo Day for finalists.
- Greater visibility with sponsors, judges, funders, and ecosystem partners.
- Potential pilot or deployment conversations.
- Additional Sahara API credits.
- Public recognition through Intron and partner channels.
Partners and supporters include organisations across tech, research, finance, health, and language access (examples mentioned in coverage include NVIDIA, Masakhane, CLEAR Global, Wema Bank, PATH, Gooey.ai, CcHUB, Turn.io, and others).
How to Apply
Interested Applicants can Register / submit Expression of Interest via the linked form on the site using the Link provided 👉 Here
If you’re an African AI builder (or part of the diaspora) working on speech, agents, healthtech, fintech, edtech, civic tech, or related fields, this is a high-signal opportunity. Register, form a team if needed, and start building.
Build voice AI for how Africa actually speaks. The tools, datasets, and prize incentives are on the table. The real prize is better, more inclusive voice technology that works in the messy, multilingual reality of African life.
