Kinyarwanda ASR: 3rd Place Solution 🏆

This repository contains the code and methodology for the 3rd place winning solution in the Digital Umuganda Kinyarwanda ASR Challenge Kaggle competition. The primary goal was to develop a high-accuracy Automatic Speech Recognition (ASR) model for the Kinyarwanda language using 1,000 hours of transcribed speech.

🚀 Methodology and Experiments

To achieve the best possible performance, we explored several state-of-the-art ASR architectures. Our approach was highly experimental, involving training and fine-tuning multiple models to compare their effectiveness on the Kinyarwanda dataset.

Models Explored

• Whisper: We fine-tuned various sizes of OpenAI's Whisper model, which provides a strong baseline for many languages (using the default tokenizer and a customized tokenizer)or Kinyarwanda).
• Conformer: We experimented with Conformer-based architectures, known for their excellent ability to capture both local and global features in audio.
• Parakeet (Winning Model): After extensive evaluation, the NVIDIA Parakeet (CTM) model family delivered the CombinedError = 0.4 × WER + 0.6 × CER Score = (1 – CombinedError) × 100, and became the foundation of our final submission. The code and notebooks related to this model can be found in the model_Parakeet/ directory.

Language Model Integration (Attempted)

We also attempted to further improve the model's accuracy by incorporating a language model (LM) for post-processing.

• KenLM & Beam Search: We trained a 5-gram KenLM model on a custom Kinyarwanda text corpus. The goal was to use this LM with a beam search decoder to refine the model's raw transcriptions and correct common grammatical errors.
• Challenges: Unfortunately, we encountered significant technical challenges and persistent errors during the integration phase, particularly with the decoding libraries (pyctcdecode). Due to these unresolved issues, we were unable to successfully incorporate the KenLM beam search decoder into our final pipeline. This remains a promising area for future improvement.

---

🛠️ Setup and Installation

To reproduce our environment and run the code, please follow these steps.

1. Clone the repository:

   git clone [https://github.com/itsazzaosman/DIgital-Umuganda-Kinyarwanda-ASR-3rd-Place-Solution.git](https://github.com/itsazzaosman/DIgital-Umuganda-Kinyarwanda-ASR-3rd-Place-Solution.git)
   cd DIgital-Umuganda-Kinyarwanda-ASR-3rd-Place-Solution

2. Create the Conda environment: The required environment files are located in the envs/ directory. Use the environment file corresponding to the Parakeet model to replicate our winning setup.

   # Example using the Parakeet environment file
   conda env create -f envs/parakeet_env.yml 
   conda activate parakeet_env

---

💡 Future Work

• Resolve KenLM Integration: The most immediate next step would be to debug the beam search decoder integration. Successfully applying the trained KenLM could significantly reduce the WER.
• Model Ensembling: Combining the predictions from the top-performing models (e.g., Parakeet and Whisper) could yield further performance gains.