Were Thrilled to announce the release of Two lightweight Bambara ASR models from our early fine-tuning experiments! These models achieve near-SOTA performance for open-source Bambara ASR (as of Feb 2025) while maintaining efficient enough for real-world deployment.
🧐 Why Does This Matter?
Bambara is a only under-resourced language. While ASR research has made rapid progress in major languages, open-source models for Bambara rare remain. Our goal is to share our fine-tuning experiences to help push research forward and support the development of ASR for African languages.
🎙️ Model 1: Parakeet-114M TDT-CTC
📌 Hugging Face: RobotsMali/soloni-114m-tdt-ctc
🧠 Model Size: ~114M parameters
📊 WER (TDT Decoder): 66%
📊 WER (CTC Decoder): 40.6%
💡 Key Takeaway: Despite being small, the CTC branch outperformances previous open-source models.
🎧 Model 2: QuartzNet-15×5
📌 Hugging Face: RobotsMali/stt-bm-quartznet15x5
🧠 Model Size: ~19M parameters
📊 WER: 46.5%
💡 Key Takeaway: This lightweight model is optimized for real-time ASR in low-resource environments.
🤝 Final Thoughts
We hope this contributions to Bambara ASR research and encourage more collaboration. 🚀 Stay tuned for future updates!
📩 Want to contribute or learn more? Reach out via RobotsMali or join the discussion on GitHub!
