Dataset

• Collected via crowd-sourcing: image prompts with voice responses (10–30 seconds).
  
  

• Includes audio files, metadata (speaker ID, age, gender, location), and partial transcripts.
  
  

• Covers five key domains: Health, Education, Agriculture, Government, Finance.
  
  

Prizes

Eligibility

• Open to researchers, students, startups, and hobbyists
  
  

• Team size: Up to 5 members
  
  

• Must use only permitted data and submit open-source code

Submission Requirements

1. submission.zip containing:
   
   

• Transcripts of the test set
  
  

• Public GitHub repository link with code, training scripts, models
  
  

• Technical report or blog post (PDF or link)
  
  

2. Track-specific experiment logs (e.g., Weights & Biases, TensorBoard)
   
   

3. Data sources declaration file (for Track C)
   
   

Evaluation Metrics

• Word Error Rate (WER) and Character Error Rate (CER)
  
  

• Combined Score formula:
  Score = (1 - (0.4 × WER + 0.6 × CER)) × 100
  
  

• Automated leaderboard based on test data
  
  

Rules & Guidelines

General Rules (All Tracks)

• No manual transcription or human correction on test data
  
  

• Code must be open-sourced under Apache-2.0, MIT, BSD-3-Clause, or MPL-2.0
  
  

• Only one leaderboard account per team
  
  

• Submissions must be reproducible (logs, scripts, configs required)
  
  

Track-Specific Notes

Track A – Small

• Use only the 540-hour transcribed dataset
  
  

• No external or unlabeled Kinyarwanda data allowed
  
  

• Cross-lingual pre-trained models allowed; fine-tuning must only use Track A data
  
  

Track B – Medium

• Use only the 1180-hour transcribed dataset
  
  

• Same rules as Track A, with a larger dataset
  
  

• Recommended GPU budget: less than or equal to 300 hours
  
  

Track C – Large

• Use 1180-hour labeled + 1170-hour unlabeled dataset
  
  

• Semi/self-supervised learning is allowed and encouraged
  
  

• Additional public speech data (e.g., Common Voice) permitted
  
  

• Must disclose all extra datasets and compute usage

Recommended Background

Required

• Experience with ASR model development
  
  

• Familiarity with Python, PyTorch or TensorFlow
  
  

• Understanding of model training and evaluation
  
  

Good to Have

• Experience with models like wav2vec 2.0, Whisper
  
  

• Knowledge of data augmentation for speech
  
  

• Experience training on large datasets
  
  

• Background in Kinyarwanda linguistics or related NLP fields

License

• Dataset: CC BY 4.0
  
  

• Participant code: Must use a permissive open-source license

Track Registration Links

• Track A – Join Track A
  
  

• Track B – Join Track B
  
  

• Track C – Join Track C