evie-8/kinyarwanda-speech-hackathon
Viewer
•
Updated
•
641k
•
260
Kinyarwanda Hackathon
ASR models and dataset for Kinyarwanda to explore how much data is needed to train accurate low-resource speech models.
-
akera/whisper-large-v3-kin-1h-v2
2B • Updated • 53Note Whisper-large-v3 fine-tuned on 1 hour of Kinyarwanda audio. Achieved 47.63% Word Error Rate (WER) and 16.97% Character Error Rate (CER).
akera/whisper-large-v3-kin-50h-v2
2B • Updated • 3Note Whisper-large-v3 trained on 50 hours of Kinyarwanda audio. Achieved 12.51% WER and 3.31% CER.
akera/whisper-large-v3-kin-100h-v2
2B • Updated • 4Note Whisper-large-v3 trained on 100 hours of Kinyarwanda audio. Achieved 10.90% WER and 2.84% CER.
akera/whisper-large-v3-kin-150h-v2
2B • Updated • 4Note Whisper-large-v3 trained on 150 hours of Kinyarwanda audio. Achieved 10.21% WER and 2.64% CER.
akera/whisper-large-v3-kin-200h-v2
2B • Updated • 3Note Whisper-large-v3 trained on 200 hours of Kinyarwanda audio. Achieved 9.82% WER and 2.56% CER.
akera/whisper-large-v3-kin-500h-v2
2B • Updated • 3Note Whisper-large-v3 trained on 500 hours of Kinyarwanda audio. Achieved 8.24% WER and 2.15% CER.
akera/whisper-large-v3-kin-1000h-v2
2B • Updated • 28Note Whisper-large-v3 trained on 1,000 hours of Kinyarwanda audio. Achieved 7.65% WER and 1.98% CER.
akera/whisper-large-v3-kin-full
2B • Updated • 14Note Whisper-large-v3 trained on the full Kinyarwanda dataset (~1400 hours). Achieved 7.14% WER and 1.88% CER.