OLMoASR is an open-source speech recognition project from the Allen Institute for AI, hosted on GitHub with about 490 stars under the MIT license and updates in 2025. Following AI2’s open-science approach, the repository documents the full path to training robust automatic speech recognition models, from raw data to evaluation.
Rather than shipping only model weights, OLMoASR shows its working: how the training data was collected, processed, and filtered, how models were trained, and how they are evaluated, for both short-form and long-form speech recognition. The training corpus, OLMoASR-Pool, is published on Hugging Face as audio-transcript pairs that anyone can download and organize using the documented structure. The result is a reproducible recipe for Whisper-style ASR that researchers can inspect, rerun, and adapt, together with the resulting open models.
The emphasis on openness is practical as well as scientific. Publishing the data pipeline alongside the models means other teams can audit what the model learned from, reproduce the training results, or substitute their own audio to build domain-specific recognizers, for example for accented, noisy, or specialized speech. Example workflows walk through setup, data preparation, training, and evaluation in sequence, so the repository doubles as a tutorial for training ASR systems end to end.
OLMoASR is free and open source under the MIT license. The code, the dataset, and the models are all publicly available; you only provide your own compute for training or fine-tuning.
It suits speech researchers and ML engineers who want to train or fine-tune their own robust ASR models with a transparent recipe, or who need to understand exactly what data and steps sit behind an open speech recognition model.