## [VHASR: A Multimodal Speech Recognition System With Vision Hotwords](https://arxiv.org/abs/2410.00822)

This repository provides the VHASR trained on Flickr8k, ADE20k, COCO, and OpenImages.

Our paper is available at https://arxiv.org/abs/2410.00822.

Our code is available at https://github.com/193746/VHASR/tree/main.

For specific details about training and testing, please refer to https://github.com/193746/VHASR/tree/main.

### Infer

If you are interested in our work, you can use large-scale data to train your own model and perform inference using the following command. Note that you should place the config file of clip in '{model_file}/clip_config' like the four pretrained models we provide. 

```sh
cd VHASR
CUDA_VISIBLE_DEVICES=1 python src/infer.py \
--model_name "{path_to_model_folder}" \
--speech_path "{path_to_speech}" \
--image_path "{path_to_image}" \
--merge_method 3
```

### Citation
```sh
@misc{hu2024vhasrmultimodalspeechrecognition,
      title={VHASR: A Multimodal Speech Recognition System With Vision Hotwords}, 
      author={Jiliang Hu and Zuchao Li and Ping Wang and Haojun Ai and Lefei Zhang and Hai Zhao},
      year={2024},
      eprint={2410.00822},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2410.00822}, 
}
```

### License: cc-by-nc-4.0