error-align 0.1.0b3

Text-to-text alignment algorithm for speech recognition error analysis.

Text-to-text alignment algorithm for speech recognition error analysis. ErrorAlign helps you dig deeper into your speech recognition projects by accurately aligning each word in a reference transcript with the model-generated transcript. Unlike traditional methods, such as Levenshtein-based alignment, it is not restricted to simple one-to-one alignment, but can map a single reference word to multiple words or subwords in the model output. This enables quick and reliable identification of error patterns in rare words, names, or domain-specific terms that matter most for your application.

Contents | Installation | Quickstart | Work-in-Progress | Citation and Research |

Installation

pip install error-align

Quickstart

from error_align import error_align

ref = "Some things are worth noting!"
hyp = "Something worth nothing period?"

alignments = error_align(ref, hyp)

Resulting alignments:

Alignment(SUBSTITUTE: "Some"- -> "Some"),
Alignment(SUBSTITUTE: -"thing" -> "things"),
Alignment(DELETE: "are"),
Alignment(MATCH: "worth" == "worth"),
Alignment(SUBSTITUTE: "noting" -> "nothing"),
Alignment(INSERT: "period")

Work-in-Progress

• Optimization for longform text.
• Efficient word-level first-pass.
• C++ version with Python bindings.

Citation and Research

@article{corticph2021alignment,
  title={A Text-To-Text Alignment Algorithm for Better Evaluation of Modern Speech Recognition Systems},
  author={corticph, Lasse and Havtorn, Jakob and Igel, Christian and Maal{\o}e, Lars and Tan, Zheng-Hua},
  journal={arXiv preprint arXiv:2509.24478},
  year={2025}
}

To reproduce results from the paper:

• Install with extra evaluation dependencies - only supported with Python 3.12:
  • pip install error-align[evaluation]
• Clone this repository:
  • git clone https://github.com/corticph/error-align.git
• Navigate to the evaluation directory:
  • cd error-align/evaluation
• Transcribe a dataset for evaluation. For example:
  • python transcribe_dataset.py --model_name whisper --dataset_name commonvoice --language_code fr
• Run evaluation script on the output file. For example:
  • python evaluate_dataset.py --transcript_file transcribed_data/whisper_commonvoice_test_fr.parquet

Notes:

• To reproduce results on the primock57 dataset, first run: python prepare_primock57.py.
• Use the --help flag to see all available options for transcribe_dataset.py and evaluate_dataset.py.
• All results reported in the paper are based on the test sets.

Collaborators:

---