You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

πŸ“š Kinyarwanda ASR Dataset

This dataset contains transcribed Kinyarwanda audio, designed to support training and evaluation of Automatic Speech Recognition (ASR) systems. It is part of a study on how varying training data volumes affect model performance using Whisper-large-v3.

πŸ“‚ Data Overview

The full dataset consists of approximately 263,000 audio samples covering 5 key domains:

  • πŸ₯ Health
  • πŸ›οΈ Government
  • πŸ’° Financial Services
  • πŸŽ“ Education
  • 🌾 Agriculture

To enable analysis of scale effects, the dataset is available in subsets of different durations:

  • audio_1h, audio_50h, audio_100h, audio_150h, audio_200h, audio_500h, audio_1000h, and train_cleaned (~1400h)

Each sample includes:

  • id: a unique identifier for the sample
  • audio: the audio recording
  • audio_language: the language code (kin for Kinyarwanda)
  • text: the corresponding transcription in Kinyarwanda
  • prompt: the domain or context of the audio (e.g., Health, Education, etc.)
  • duration: the length of the audio clip in seconds
  • speaker_id: an anonymized identifier for the speaker

πŸ› οΈ Usage Example

You can load and explore the dataset using the πŸ€— Datasets library:

from datasets import load_dataset

# Load a specific subset, e.g., 100 hours
dataset = load_dataset("evie-8/kinyarwanda-speech-hackathon", "audio_100h")
Downloads last month
1,329