MMS-LLaMA: Efficient LLM-based Audio-Visual Speech Recognition with Minimal Multimodal Speech Tokens

This repository contains the PyTorch implementation of the following paper:

MMS-LLaMA: Efficient LLM-based Audio-Visual Speech Recognition with Minimal Multimodal Speech Tokens
Authors: Jeong Hun Yeo*, Hyeongseop Rha*, Se Jin Park, Yong Man Ro (*Equal contribution)
Paper Link: http://arxiv.org/abs/2503.11315

Introduction

MMS-LLaMA is an efficient multimodal speech LLM framework, for AVSR that minimizes the length of multimodal speech tokens while preserving their linguistic content.

Environment Setup

conda create -n mms-llama python=3.9 -y
conda activate mms-llama
git clone https://github.com/JeongHun0716/MMS-LLaMA
cd MMS-LLaMA

# PyTorch and related packages
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu121
pip install numpy==1.23.5 scipy opencv-python
pip install editdistance python_speech_features einops soundfile sentencepiece tqdm tensorboard unidecode librosa
pip install omegaconf==2.0.6 hydra-core==1.0.7 (If your pip version > 24.1, please run "python3 -m pip install --upgrade pip==24.0")
pip install transformers==4.47.1 peft==0.14.0 bitsandbytes==0.45.0
cd fairseq
pip install --editable ./

Preparation

Before running any training or evaluation, you must update the dataset file paths in the tsv files. These tsv files contain placeholders (e.g., {LRS3_ROOT}) that need to be replaced with the absolute paths to your local copies of the datasets. The provided script (update_dataset_paths.py) automates this process, ensuring that all references in the tsv files point to the correct locations on your system.

The required datasets are:

• VoxCeleb2
  Download from the VoxCeleb2 website.

• LRS3
  Download from the LRS3-TED dataset page.

Once you have downloaded these datasets, you should pre-process every video clip to crop the mouth regions. You can follow the pre-processing instructions provided in Auto-AVSR.

Note that for the LRS3 and VoxCeleb2 datasets, the facial landmarks are already provided in the Auto-AVSR repository.

After the pre-processing, update the tsv files with the absolute paths to the dataset directories using the provided script. This ensures that all dataset references point to the correct locations on your system.

python update_dataset_paths.py --input_dir ./ --vox2 'path for the VoxCeleb2 dataset' --lrs3 'path for the LRS3 dataset' 

For example:

python update_dataset_paths.py --input_dir ./ --vox2 /Dataset/vox2 --lrs3 /Dataset/lrs3

The above command updates the placeholder paths in the tsv files to your absolute dataset paths.

Each tsv files contains one line per data sample, with the following fields separated by a tab (\t):

• used dataset
• video_path
• audio_path
• num_video_frames
• num_audio_frames
• speech_rate

Below are the expected directory structures for LRS3 dataset:

LRS3

lrs3/
├── lrs3_video_seg24s/              
│   ├── pretrain/
│   ├── test/
│   ├── trainval/            
└── lrs3_text_seg24s/
    ├── pretrain/
    ├── test/
    └── trainval/    

Training

Train a new MMS-LLaMA

bash scripts/train.sh

Note: For training on an 8 GPU RTX 3090 setup, the 433h model can be trained in approximately 6 hours, while the 1759h model requires about 20 hours.

Evaluation of the MMS-LLaMA

To evaluate the performance of MMS-LLaMA, execute the evaluation script by running:

bash scripts/eval.sh

Evaluation of the MMS-LLaMA, under noisy environment

To evaluate the performance of MMS-LLaMA in a noisy environment, run the evaluation script using:

bash scripts/eval_snr.sh

Pretrained Models

1. Download the AV-HuBERT Large model from this link
2. Download the Whisper-medium.en from this link
3. Download the LLaMA-3.2 3B model from this link

After downloading, make sure to place the models in the correct directories:

• The large_vox_iter5.pt(AV-HuBERT) model should be placed in the pretrained_models/avhubert folder.
• Place the 433h checkpoint in the pretrained_models/mms-llama/433h folder.
• Place the 1759h checkpoint in the pretrained_models/mms-llama/1759h folder.
• Place the speech rate predictor checkpoint in the pretrained_models/sr_predictor folder.

MMS-LLaMA

Model	Used Datasets	Training data (# hours)	WER(%), Clean	WER(%), Noisy
ckpt.pt	LRS3	433	0.90	2.4
ckpt.pt	LRS3, VoxCeleb2	1759	0.72	1.9

You can download the pre-trained models using wget with the following command:

# 433h model
wget -O ckpt.pt "https://www.dropbox.com/scl/fi/uiaxa2lgjze4mt7tdi5wu/checkpoint_best.pt?rlkey=o62sc6ann8xm3gpkyj4yk3rwe&st=s5q385op&dl=1"

# 1759h model
wget -O ckpt.pt "https://www.dropbox.com/scl/fi/ou28xe2k9ampxsf4ihoft/checkpoint_best.pt?rlkey=a4q1qgigodhrgwqi9lgsalj7f&st=ga8z79vc&dl=1"

Speech Rate Predictor

Model	Used Datasets	Training data (# hours)
ckpt.pt	LRS3	433

wget -O ckpt.pt "https://www.dropbox.com/scl/fi/rc6jmbzdvj8afn84z47qt/checkpoint.pt?rlkey=aoa0ifkdydgm9gjmt2ljwpgrc&st=we9qoqtb&dl=1"

Citation

If you find this work useful in your research, please cite the paper:

@article{yeo2025mms,
  title={MMS-LLaMA: Efficient LLM-based Audio-Visual Speech Recognition with Minimal Multimodal Speech Tokens},
  author={Yeo, Jeong Hun and Rha, Hyeongseop and Park, Se Jin and Ro, Yong Man},
  journal={arXiv preprint arXiv:2503.11315},
  year={2025}
}

Acknowledgement

This project is based on the avhubert, auto-avsr, and fairseq code. We would like to thank the developers of these projects for their contributions and the open-source community for making this work possible.