Kinyawhisper

🚀 KinyaWhisper

KinyaWhisper is a fine - tuned OpenAI's Whisper model for Kinyarwanda automatic speech recognition, offering a reproducible baseline for low - resource indigenous languages.

🚀 Quick Start

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 Examples

Basic Usage

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)

🔧 Technical Details

⚠️ Important Note

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.

📚 Documentation

Model Index

• Name: KinyaWhisper
  • Results:
    • Task:
      • Name: Automatic Speech Recognition
      • Type: automatic - speech - recognition
    • Dataset:
      • Name: KinyaWhisper Custom Dataset
      • Type: custom
      • Config: kinyarwanda
    • Metrics:
      • Name: WER
      • Type: wer
      • Value: 51.85

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}
}

📄 License

This project is licensed under the MIT license.

📬 Contact

Maintained by Gabriel Baziramwabo. ✉️ gabriel@benax.rw 🔗 https://benax.rw