README.md

metadata

title: Fonbench Leaderboard
emoji: 🌍
colorFrom: blue
colorTo: gray
sdk: gradio
sdk_version: 6.19.0
python_version: '3.13'
app_file: app.py
pinned: false
license: mit
short_description: ASR benchmark for Fon (low-resource African tonal language)

🎤 FonBench — ASR Benchmark and Adaptation for Fon

Interactive leaderboard comparing six pre-trained multilingual ASR models on the Fon language (low-resource tonal language of Benin, ~8M speakers), and four adaptation strategies (LoRA and full fine-tuning) on the best model.

📊 Phase 1 — Zero-shot evaluation

Six state-of-the-art multilingual ASR models evaluated without adaptation: MMS-1b, XLSR-53, OmniASR-CTC, AfriHuBERT, SeamlessM4T, Whisper-small.

🏆 Phase 2 — Adaptation

Two adaptation strategies (LoRA at ranks 8 and 32, Full Fine-Tuning at LR 1e-5 and 5e-5) applied to the best zero-shot model.

Best result: WER 19.27% (Full Fine-Tuning, LR 5e-5) — a 64-point improvement over zero-shot.

🔗 Resources

• 📦 Models on HuggingFace Hub: (coming soon)
• 💻 Source code on GitHub: (coming soon)
• 📈 Experiment tracking on W&B: FonBench project
• 📄 Dataset: alaleye/fon (CC-BY-4.0)

👤 Author

Inès Assia HOUNKPONOU — Master IA, ESGIS Bénin, 2026 Supervisor: Prof. Fréjus LALEYE

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference