Model
This repository contains the third version of our Automatic Speech Recognition and Subtitle Generation model for Flemish Dutch. Compared to the second version, the model is a fully Pytorch-based model without dependency on Kaldi-features, facilitating simple deployment and finetuning.
The model has been trained on 300 hours of verbatim annotated Flemish data from CGN (with 3-fold noise augmentation), 700 hours of Netherlands Dutch data from CGN, as well as 14000 hours of weakly-supervised subtitled Flemish broadcast media data. Additionally, we enriched the training data by generating contextualised verbatim pseudo-labels conditioned on the subtitles, to improve rare word recognition in verbatim transcripts.
The model can generate both an exact verbatim transcription with annotation tags as well as a fully formatted and cleaned up subtitle transcription. It outputs both modalities with separate decoders. The model consists of 180M parameters and requires 2-6GB GPU RAM for inference.
Version: August 2025
Usage
This repository only hosts the pre-trained model itself and the configuration files. To download this repository, follow the instructions by Huggingface.
Usage of this model and an example test file to perform inference - fully integrated with a VAD pipeline - can be found on Github. We incorporated the Silero VAD in this pipeline, which is uploaded as part of the model files.
The model is released under a Creative Commons Non-Commercial license.
Citation
If you use this model, please cite the research paper:
@article{poncelet2024,
author = "Poncelet, Jakob and Van hamme, Hugo",
title = "Leveraging Broadcast Media Subtitle Transcripts for Automatic Speech Recognition and Subtitling",
year={2024},
journal={arXiv preprint arXiv:2502.03212},
url = {https://arxiv.org/abs/2502.03212}
Contact
Jakob Poncelet: [email protected]