Deprecation Notice (June 16, 2025)

This dataset is deprecated. A newer version is available on Hugging Face Datasets, featuring higher data quality and reliability, although with fewer overall samples.

View the updated dataset on Hugging Face

ATC Dataset - Fine-Tuning Whisper

This dataset was created to fine-tune OpenAI's Whisper model for improving transcription accuracy in Air Traffic Control (ATC) communications. The dataset contains transcriptions and corresponding audio files from two main sources: ATCO2 and the UWB-ATCC corpus, specifically selected for aviation-related communications. The dataset is publicly available on Hugging Face for use in Automatic Speech Recognition (ASR) projects.

For more details on the fine-tuning process, check out the blog post and the corresponding GitHub repository.

Dataset Overview

• Dataset Name: ATC Dataset
• Total Samples: 11.9k (Training), 2.93k (Test)
• Data Sources:
  • ATCO2 Corpus (1-hour test subset)
  • UWB-ATCC Corpus
• Format: Audio files (WAV format) with corresponding transcriptions.
• The Transciption Ground Truth is within the text column.
• The Audio Ground Truth is within the audio column.
• License: MIT

This dataset is particularly useful for training speech recognition models like Whisper on short, domain-specific audio transmissions, such as those between pilots and air traffic controllers.

Key Features

• Domain-specific: Tailored to ATC communications with specialized phraseology and terms.
• Diverse accents: Contains multiple accent variations to reflect real-world international aviation communication.
• Cleaned Data: Includes only high-quality samples after filtering erroneous or incomplete transcriptions.

Usage

1. Install Dependencies:
   Use Hugging Face's datasets library to load the dataset:

   from datasets import load_dataset
   dataset = load_dataset("jacktol/atc-dataset")
   

2. Training:
   The dataset is ready for speech recognition tasks such as fine-tuning Whisper models. It includes training and test splits to evaluate models based on Word Error Rate (WER).

License

This dataset is shared under the MIT License. You are free to use, modify, and distribute it as long as you provide proper attribution.