Kinyarwanda-Health-ASR

Model Description

This model is a fine-tuned version of Wav2Vec2-BERT 2.0 for Kinyarwanda automatic speech recognition (ASR) Health. The model is has a WER of 5.95% in the health domain

• Developed by: Elvis Tata
• Model type: Speech Recognition (ASR)
• Language: Kinyarwanda
• License: CC-BY-4.0
• Finetuned from model [badrex/w2v-bert-2.0-kinyarwanda-asr]: [facebook/w2v-bert-2.0]

Model Sources [optional]

• Repository: Space
• Demo: Demo

Direct Use

from transformers import Wav2Vec2BertProcessor, Wav2Vec2BertForCTC
import torch
import torchaudio

# load the model and processor
processor = AutoProcessor.from_pretrained("ElvisTata2024/Kinyarwanda-Health-ASR")
model = AutoModelForCTC.from_pretrained("ElvisTata2024/Kinyarwanda-Health-ASR")

# load audio
audio_input, sample_rate = torchaudio.load("path/to/audio.wav")

# preprocess
inputs = processor(audio_input.squeeze(), sampling_rate=sample_rate, return_tensors="pt")

# inference
with torch.no_grad():
    logits = model(**inputs).logits

# decode
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)[0]
print(transcription)

Training Data

-DigitalUmuganda/ASR_Fellowship_Challenge_Dataset

Results

Model Architecture and Objective

• Base model: Wav2Vec2-BERT 2.0
• Architecture: transformer-based with convolutional feature extractor
• Parameters: 580M
• Objective: connectionist temporal classification (CTC),KenLM Beam search decoding

Citation

@misc{w2v_bert_kinyarwanda_asr,
  author = {Badr M. Abdullah},
  title = {Adapting Wav2Vec2-BERT 2.0 for Kinyarwanda ASR},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/badrex/w2v-bert-2.0-kinyarwanda-asr-1000h}
}
@misc{Kinyarwanda-Health-ASR,
  author = {Elvis Tata Tanghang},
  title = {Adapting Wav2Vec2-BERT 2.0 for Kinyarwanda ASR Health},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/ElvisTata2024/Kinyarwanda-Health-ASR}
}