Workshop Challenge

Challenge Winners

Mission Prizes

$100 each

• EducationJohn Fiewor — GradrAIEmpowers students who may reason and communicate better in local languages via code-switching to better understand questions and perform better on Computer-Based Tests, reducing the risk of their intelligence being overshadowed by a language barrier.
• AgricultureTriumph team — SOAT Voice agentUses code-switched voice to interact with physical agricultural irrigation systems and provide advisory details on soil pH, crop yields, and more.
• Language InclusionJitimay Josué — Dawa VoiceFound a way to build for Kirundi, a language spoken by 12 million people in Burundi that is unsupported by most AI systems.
• Female-LedEmiA voice-based respiratory triage app with WHO-grounded urgency/triage labels that are automatically extracted and auto-populate downstream forms for community health worker follow-up.
• User ExperiencePAYGO team — Koko AIA voice-based bookkeeping system that automatically adds/subtracts petty cash and tracks inventory, making bookkeeping fast, natural, and accessible during everyday business while ensuring financial data never leaves the device. It featured a better UX than the first-place solution but fell short on the thoroughness of the benchmark report.

The Challenge

Across Africa, people rarely speak in neat, single-language blocks. Conversations move naturally between English/French and languages like Swahili, Hausa, Zulu, Yoruba, Igbo, Pidgin, Amharic, or Kinyarwanda, often within the same sentence. This is code-switching, and it remains one of the most under-solved problems in voice AI: systems built for clean, monolingual speech tend to break down in the real, messy, multilingual settings where they're needed most.

The MLC (Africa) × Intron Agentic Voice AI Challenge invites participants at Deep Learning Indaba 2026 to build voice-driven agentic applications that work the way Africa actually speaks, using Intron's Sahara APIs as a foundation to help surface how well current speech models handle code-switched audio in real-world settings.

Build an agentic voice application

A solution where voice is used to drive a downstream task or action, not just transcribe speech. Examples include a voice assistant that completes a booking, a voice-driven triage tool, a spoken-language form-filler, or a voice interface that takes an action on a user's behalf. Each submission must:

1. 1Use voice to achieve a downstream task (the agentic component), not simply convert speech to text.
2. 2Benchmark performance on code-switched audio input across at least three speech models, including an Intron Sahara API, and at least two other models of the team's choice (global, local, open-source, or commercial).
3. 3Target a specific vertical or use case (see categories below).
4. 4Where permitted, submit the code-switched audio samples used for testing, with basic metadata (language pair, domain, accent/country, device type, noise conditions).

As a bonus tip, we are providing free Adaption labs credit to teams that may want to train/finetune. Here are some ideas on how to use the credit. You can check out Adaption's documentation, or ask questions on Adaption's discord.

Suggested Categories

Teams are encouraged, but not required, to align their solution with one of the following:

• →Health
• →Fintech, Telco & Customer Experience
• →Legal & Public Services
• →Agriculture & Education
• →Other high-impact use cases (humanitarian response, accessibility, media, transport, commerce, civic engagement, etc.)

What to Submit

• Solution DescriptionA short description of the problem, target users, solution, and technical overview highlighting key design decisions.
• DemoA short demo video of a working prototype or integrated application — a link to an unlisted/public YouTube video.
• DocsCode or technical documentation (integrations into proprietary applications from startups or enterprises do not need to submit confidential code).
• Benchmark reportResults comparing at least three speech models (including Sahara), with pros and cons of using each.
• Ethics / InclusionA brief note on privacy, consent, safety, and responsible data use.
• Benchmark Audios (optional)Consented, de-identified code-switched audio samples used for benchmarking in your selected category, with metadata.

Key Dates

• 3rd August 2026Registration opens
• 6th August 2026Submission deadline
• 7th August 2026Finalists announced at the workshop (Main Plenary Hall)
• Same day / later onlineWinner(s) announced

Teams that want to continue building after Indaba can carry over their submission into Intron's full Sahara CodeSwitch Africa Challenge (build period through mid-September, winners announced October 1).

Judging Criteria

• Real-world impactIs the solution agentic, and does it address a genuine user need for a meaningful population?
• Code-switching performanceQuality and fairness of the benchmark comparison across models.
• Product qualityUsability and workflow fit of the agentic application.
• Technical executionReliability, integration, latency, robustness.
• Ethics, safety & inclusionConsent, privacy, bias awareness, respect for user dignity.

Bonus consideration for: strong benchmark design, low-resource language coverage, offline/low-bandwidth readiness, and clear potential to continue into the full Sahara CodeSwitch Africa Challenge.

Provided by Intron

• →Access to Sahara v2.5 code-switching APIs across supported language pairs.
• →Access to the AfriSwitch benchmark dataset with code-switching samples in 12 language pairs.
• →A public benchmarking framework for comparing model performance.