Parulekar, Amruta, and Preethi Jyothi. 2025. “LASER: An LLM-Based ASR Scoring and Evaluation Rubric.” arXiv:2510.07437. Preprint, arXiv, October 8. https://doi.org/10.48550/arXiv.2510.07437.

Notes

• LLM is used for computing the errors at the sentence level instead of using WER as a metric
• The penalties are determined for the errors based on the type of errors. 1 for major penalty and 0.5 for minor penalty
• Sentence level score is calculated by $1-\frac{\text{Total penalty}}{\text{Number of reference words}}$
• They have also tried fine tuning an SLM for computing the errors instead of LLM since LLMs are expensive

In-text annotations