OLMoASR-Pool is a web-scale audio-text dataset collected from the public internet, consisting of approximately 3M hours of audio and 17M transcripts.

With OLMoASR-Pool, we trained OLMoASR 💬🎙️, a series of English speech recognition models and observed strong generalization and robust capabilities!

Content

• The dataset contains 18,761,823 unique IDs spanning approximately 3.4M hours of audio.
• It also spans across a variety speaking styles, accents and audio setups such as news segments 📰, podcasts 🎙️, outdoors 🌳🏙️, crowds 🧑‍🤝‍🧑, speeches 🎤, commentary 🗣️, interviews 🤳 and more!
• OLMoASR-Pool is multilingual as it can contain non-English audio/transcripts. To retrieve an English-only dataset, it is critical to perform audio-text language alignment.
• After downloading the collection for training, only 3M hours of audio and 17M transcripts remains.

Usage

1. Download from HuggingFace
   • Retrieve HF access token from here to gain access to the dataset.
   • Run pip install huggingface_hub[cli]
   • Run huggingface-cli login in your CLI and paste the HF access token to login
   • Use the code below to access the IDs

     from datasets import load_dataset
     dataset = load_dataset("allenai/OLMoASR-Pool", streaming=True)
     print(dataset) # features: ['id']
     print(next(iter(dataset['train'])))
     

   • If you're downloading all the IDs, you can run the code below

   from datasets import load_dataset
   dataset = load_dataset("allenai/OLMoASR-Pool", streaming=False, cache_dir=<where you want to download the IDs to>)
   

2. Download the audio and transcript files from ID information.
3. Preprocess the audio and transcript files. Follow the instructions at the OLMoASR repo

Uses

The collection was used to train a speech recognition model, but it can also be used in research areas such as conversational data, audio understanding, speaker diarization, voice detection and more.

License

This dataset is licensed under ODC-BY. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines.

Reference

@misc{ngo2025olmoasropenmodelsdata,
      title={OLMoASR: Open Models and Data for Training Robust Speech Recognition Models}, 
      author={Huong Ngo and Matt Deitke and Martijn Bartelds and Sarah Pratt and Josh Gardner and Matt Jordan and Ludwig Schmidt},
      year={2025},
      eprint={2508.20869},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2508.20869}, 
}