benax-rw/KinyaWhisper

🗣️ KinyaWhisper

KinyaWhisper is a fine-tuned version of OpenAI’s Whisper model for automatic speech recognition (ASR) in Kinyarwanda. It was trained on 102 manually labeled .wav files and serves as a reproducible baseline for speech recognition in low-resource, indigenous languages.

🔧 Usage

To run inference on your own audio files using the fine-tuned KinyaWhisper model:

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torchaudio

# Load fine-tuned KinyaWhisper model and processor from Hugging Face
model = WhisperForConditionalGeneration.from_pretrained("benax-rw/KinyaWhisper")
processor = WhisperProcessor.from_pretrained("benax-rw/KinyaWhisper")

# Load and preprocess audio
waveform, sample_rate = torchaudio.load("your_audio.wav")
inputs = processor(waveform.squeeze(), sampling_rate=sample_rate, return_tensors="pt")

# Generate prediction
predicted_ids = model.generate(inputs["input_features"])
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]

print("🗣️ Transcription:", transcription)

🏋️ Taining Details

• Model: openai/whisper-small • Epochs: 80 • Batch size: 4 • Learning rate: 1e-5 • Optimizer: Adam • Final loss: 0.00024 • WER: 51.85%

⚠️Limitations

The model was trained on a small dataset (102 samples). It performs best on short, clear Kinyarwanda utterances and may struggle with longer or noisy audio. This is an early-stage educational model, not yet suitable for production use.

📚 Citation

If you use this model, please cite:

@misc{baziramwabo2025kinyawhisper,
  author       = {Gabriel Baziramwabo},
  title        = {KinyaWhisper: Fine-Tuning Whisper for Kinyarwanda ASR},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/benax-rw/KinyaWhisper}},
  note         = {Version 1.0}
}

📬 Contact

Maintained by Gabriel Baziramwabo. ✉️ [email protected] 🔗 https://benax.rw