Use Kaldi alignment steps in https://www.eleanorchodroff.com/tutorial/kaldi/forced-alignment.html to generate ctm of words in the reference.
Prepare the dataset in this format:
audio_file_name, reference, start_time_of_words (separated by space), end_time_of_words (separated by space)
For example,
kb_all/kb_data_clean_m4a/hindi/test/audio/844424932919645-1202-f.m4a सुधीर मिश्रा की ये फिल्म लेखक शरत चंद्र चट्टोपाध्याय के उपन्यास देवदास पर आधारित होगी 0.53 0.95 1.23 1.43 1.52 1.77 2.26 2.62 3.28 3.87 3.96 4.39 4.79 4.91 5.41 0.95 1.23 1.43 1.52 1.77 2.26 2.62 3.04 3.84 3.96 4.39 4.76 4.91 5.37 5.55
https://zenodo.org/records/11162885
If you find this useful, please cite as
@article{ravi2024teles,
title={TeLeS: Temporal Lexeme Similarity Score to Estimate Confidence in End-to-End ASR},
author={Ravi, Nagarathna, T, Thishyan Raj and Arora, Vipul},
journal={arXiv preprint arXiv:2401.03251},
year={2024}
}