SubSONAR evaluates the quality of SRT files using the multilingual multimodal SONAR model.

The evaluation accounts for the semantic similarity (computed as a cosine similarity) between each subtitle block and the corresponding audio to which the block is assigned to (through the timestamps in the SRT). The returned scores range in [-1, 1] where the higher, the better.

SubSONAR is available both in its open source repository on GitHub and on PyPi.


SubSONAR is licensed under Apache Version 2.0. However, the SONAR encoders have a dedicated license that can be found in their repository LICENSE. Please check the license for the encoders you are using.


If using this repository, please consider citing:

title = {{SBAAM! Eliminating Transcript Dependency in Automatic Subtitling}},
author = {Gaido, Marco and Papi, Sara and Negri, Matteo and Cettolo, Mauro and Bentivogli, Luisa},
booktitle = "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
year = "2024",
address = "Bangkok, Thailand",