Hugging Face's logo

Datasets:
nguyenvulebinh
/
AVCocktail

Modalities:
Text
Formats:
webdataset
Size:
1K - 10K
ArXiv:
Libraries:
Datasets
WebDataset
License:
Dataset card Data Studio Files
xet
Community
1
AVCocktail / README.md
nguyenvulebinh's picture
nguyenvulebinh
Update README.md
b189d12 verified
preview code
|
raw
history blame contribute delete
26.8 kB
 ---
license: cc-by-sa-4.0
configs:
- config_name: video_0
data_files:
- split: asd_chunk
path: "videos/video_0/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_0/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_0/gold-chunk-eval-0.tar"
- config_name: video_1
data_files:
- split: asd_chunk
path: "videos/video_1/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_1/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_1/gold-chunk-eval-0.tar"
- config_name: video_2
data_files:
- split: asd_chunk
path: "videos/video_2/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_2/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_2/gold-chunk-eval-0.tar"
- config_name: video_3
data_files:
- split: asd_chunk
path: "videos/video_3/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_3/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_3/gold-chunk-eval-0.tar"
- config_name: video_4
data_files:
- split: asd_chunk
path: "videos/video_4/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_4/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_4/gold-chunk-eval-0.tar"
- config_name: video_5
data_files:
- split: asd_chunk
path: "videos/video_5/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_5/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_5/gold-chunk-eval-0.tar"
- config_name: video_6
data_files:
- split: asd_chunk
path: "videos/video_6/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_6/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_6/gold-chunk-eval-0.tar"
- config_name: video_7
data_files:
- split: asd_chunk
path: "videos/video_7/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_7/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_7/gold-chunk-eval-0.tar"
- config_name: video_8
data_files:
- split: asd_chunk
path: "videos/video_8/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_8/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_8/gold-chunk-eval-0.tar"
- config_name: video_9
data_files:
- split: asd_chunk
path: "videos/video_9/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_9/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_9/gold-chunk-eval-0.tar"
- config_name: video_10
data_files:
- split: asd_chunk
path: "videos/video_10/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_10/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_10/gold-chunk-eval-0.tar"
- config_name: video_11
data_files:
- split: asd_chunk
path: "videos/video_11/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_11/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_11/gold-chunk-eval-0.tar"
- config_name: video_12
data_files:
- split: asd_chunk
path: "videos/video_12/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_12/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_12/gold-chunk-eval-0.tar"
- config_name: video_13
data_files:
- split: asd_chunk
path: "videos/video_13/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_13/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_13/gold-chunk-eval-0.tar"
- config_name: video_14
data_files:
- split: asd_chunk
path: "videos/video_14/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_14/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_14/gold-chunk-eval-0.tar"
- config_name: video_15
data_files:
- split: asd_chunk
path: "videos/video_15/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_15/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_15/gold-chunk-eval-0.tar"
- config_name: video_16
data_files:
- split: asd_chunk
path: "videos/video_16/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_16/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_16/gold-chunk-eval-0.tar"
- config_name: video_17
data_files:
- split: asd_chunk
path: "videos/video_17/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_17/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_17/gold-chunk-eval-0.tar"
- config_name: video_18
data_files:
- split: asd_chunk
path: "videos/video_18/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_18/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_18/gold-chunk-eval-0.tar"
- config_name: video_19
data_files:
- split: asd_chunk
path: "videos/video_19/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_19/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_19/gold-chunk-eval-0.tar"
- config_name: video_20
data_files:
- split: asd_chunk
path: "videos/video_20/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_20/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_20/gold-chunk-eval-0.tar"
- config_name: video_21
data_files:
- split: asd_chunk
path: "videos/video_21/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_21/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_21/gold-chunk-eval-0.tar"
- config_name: video_22
data_files:
- split: asd_chunk
path: "videos/video_22/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_22/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_22/gold-chunk-eval-0.tar"
- config_name: video_23
data_files:
- split: asd_chunk
path: "videos/video_23/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_23/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_23/gold-chunk-eval-0.tar"
- config_name: video_24
data_files:
- split: asd_chunk
path: "videos/video_24/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_24/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_24/gold-chunk-eval-0.tar"
- config_name: video_25
data_files:
- split: asd_chunk
path: "videos/video_25/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_25/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_25/gold-chunk-eval-0.tar"
- config_name: video_26
data_files:
- split: asd_chunk
path: "videos/video_26/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_26/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_26/gold-chunk-eval-0.tar"
- config_name: video_27
data_files:
- split: asd_chunk
path: "videos/video_27/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_27/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_27/gold-chunk-eval-0.tar"
- config_name: video_28
data_files:
- split: asd_chunk
path: "videos/video_28/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_28/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_28/gold-chunk-eval-0.tar"
- config_name: video_29
data_files:
- split: asd_chunk
path: "videos/video_29/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_29/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_29/gold-chunk-eval-0.tar"
- config_name: video_30
data_files:
- split: asd_chunk
path: "videos/video_30/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_30/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_30/gold-chunk-eval-0.tar"
- config_name: video_31
data_files:
- split: asd_chunk
path: "videos/video_31/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_31/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_31/gold-chunk-eval-0.tar"
- config_name: video_32
data_files:
- split: asd_chunk
path: "videos/video_32/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_32/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_32/gold-chunk-eval-0.tar"
- config_name: video_33
data_files:
- split: asd_chunk
path: "videos/video_33/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_33/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_33/gold-chunk-eval-0.tar"
- config_name: video_34
data_files:
- split: asd_chunk
path: "videos/video_34/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_34/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_34/gold-chunk-eval-0.tar"
- config_name: video_35
data_files:
- split: asd_chunk
path: "videos/video_35/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_35/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_35/gold-chunk-eval-0.tar"
- config_name: video_36
data_files:
- split: asd_chunk
path: "videos/video_36/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_36/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_36/gold-chunk-eval-0.tar"
- config_name: video_37
data_files:
- split: asd_chunk
path: "videos/video_37/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_37/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_37/gold-chunk-eval-0.tar"
- config_name: video_38
data_files:
- split: asd_chunk
path: "videos/video_38/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_38/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_38/gold-chunk-eval-0.tar"
- config_name: video_39
data_files:
- split: asd_chunk
path: "videos/video_39/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_39/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_39/gold-chunk-eval-0.tar"
- config_name: video_40
data_files:
- split: asd_chunk
path: "videos/video_40/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_40/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_40/gold-chunk-eval-0.tar"
- config_name: video_41
data_files:
- split: asd_chunk
path: "videos/video_41/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_41/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_41/gold-chunk-eval-0.tar"
- config_name: video_42
data_files:
- split: asd_chunk
path: "videos/video_42/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_42/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_42/gold-chunk-eval-0.tar"
- config_name: video_43
data_files:
- split: asd_chunk
path: "videos/video_43/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_43/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_43/gold-chunk-eval-0.tar"
- config_name: video_44
data_files:
- split: asd_chunk
path: "videos/video_44/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_44/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_44/gold-chunk-eval-0.tar"
- config_name: video_45
data_files:
- split: asd_chunk
path: "videos/video_45/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_45/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_45/gold-chunk-eval-0.tar"
- config_name: video_46
data_files:
- split: asd_chunk
path: "videos/video_46/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_46/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_46/gold-chunk-eval-0.tar"
- config_name: video_47
data_files:
- split: asd_chunk
path: "videos/video_47/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_47/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_47/gold-chunk-eval-0.tar"
- config_name: video_48
data_files:
- split: asd_chunk
path: "videos/video_48/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_48/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_48/gold-chunk-eval-0.tar"
- config_name: video_49
data_files:
- split: asd_chunk
path: "videos/video_49/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_49/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_49/gold-chunk-eval-0.tar"
- config_name: video_50
data_files:
- split: asd_chunk
path: "videos/video_50/asd-chunk-eval-0.tar"
- split: fixed_chunk
path: "videos/video_50/fixed-chunk-eval-0.tar"
- split: gold_chunk
path: "videos/video_50/gold-chunk-eval-0.tar"
- config_name: labels
data_files:
- split: video_0
path: "labels/video_0/label-eval-0.tar"
- split: video_1
path: "labels/video_1/label-eval-0.tar"
- split: video_2
path: "labels/video_2/label-eval-0.tar"
- split: video_3
path: "labels/video_3/label-eval-0.tar"
- split: video_4
path: "labels/video_4/label-eval-0.tar"
- split: video_5
path: "labels/video_5/label-eval-0.tar"
- split: video_6
path: "labels/video_6/label-eval-0.tar"
- split: video_7
path: "labels/video_7/label-eval-0.tar"
- split: video_8
path: "labels/video_8/label-eval-0.tar"
- split: video_9
path: "labels/video_9/label-eval-0.tar"
- split: video_10
path: "labels/video_10/label-eval-0.tar"
- split: video_11
path: "labels/video_11/label-eval-0.tar"
- split: video_12
path: "labels/video_12/label-eval-0.tar"
- split: video_13
path: "labels/video_13/label-eval-0.tar"
- split: video_14
path: "labels/video_14/label-eval-0.tar"
- split: video_15
path: "labels/video_15/label-eval-0.tar"
- split: video_16
path: "labels/video_16/label-eval-0.tar"
- split: video_17
path: "labels/video_17/label-eval-0.tar"
- split: video_18
path: "labels/video_18/label-eval-0.tar"
- split: video_19
path: "labels/video_19/label-eval-0.tar"
- split: video_20
path: "labels/video_20/label-eval-0.tar"
- split: video_21
path: "labels/video_21/label-eval-0.tar"
- split: video_22
path: "labels/video_22/label-eval-0.tar"
- split: video_23
path: "labels/video_23/label-eval-0.tar"
- split: video_24
path: "labels/video_24/label-eval-0.tar"
- split: video_25
path: "labels/video_25/label-eval-0.tar"
- split: video_26
path: "labels/video_26/label-eval-0.tar"
- split: video_27
path: "labels/video_27/label-eval-0.tar"
- split: video_28
path: "labels/video_28/label-eval-0.tar"
- split: video_29
path: "labels/video_29/label-eval-0.tar"
- split: video_30
path: "labels/video_30/label-eval-0.tar"
- split: video_31
path: "labels/video_31/label-eval-0.tar"
- split: video_32
path: "labels/video_32/label-eval-0.tar"
- split: video_33
path: "labels/video_33/label-eval-0.tar"
- split: video_34
path: "labels/video_34/label-eval-0.tar"
- split: video_35
path: "labels/video_35/label-eval-0.tar"
- split: video_36
path: "labels/video_36/label-eval-0.tar"
- split: video_37
path: "labels/video_37/label-eval-0.tar"
- split: video_38
path: "labels/video_38/label-eval-0.tar"
- split: video_39
path: "labels/video_39/label-eval-0.tar"
- split: video_40
path: "labels/video_40/label-eval-0.tar"
- split: video_41
path: "labels/video_41/label-eval-0.tar"
- split: video_42
path: "labels/video_42/label-eval-0.tar"
- split: video_43
path: "labels/video_43/label-eval-0.tar"
- split: video_44
path: "labels/video_44/label-eval-0.tar"
- split: video_45
path: "labels/video_45/label-eval-0.tar"
- split: video_46
path: "labels/video_46/label-eval-0.tar"
- split: video_47
path: "labels/video_47/label-eval-0.tar"
- split: video_48
path: "labels/video_48/label-eval-0.tar"
- split: video_49
path: "labels/video_49/label-eval-0.tar"
- split: video_50
path: "labels/video_50/label-eval-0.tar"
---
# AVSRCocktail: Audio-Visual Speech Recognition for Cocktail Party Scenarios
**Official implementation** of "[Cocktail-Party Audio-Visual Speech Recognition](https://arxiv.org/abs/2506.02178)" (Interspeech 2025).
A robust audio-visual speech recognition system designed for multi-speaker environments and noisy cocktail party scenarios. The model combines lip reading and audio processing to achieve superior performance in challenging acoustic conditions with background noise and speaker interference.
## Getting Started
### Sections
1. <a href="#install">Installation</a>
2. <a href="#evaluation">Evaluation</a>
3. <a href="#training">Training</a>
## <a id="install">1. Installation </a>
Following this steps:
```sh
# Clone the baseline code repo
git clone https://github.com/nguyenvulebinh/AVSRCocktail.git
cd AVSRCocktail
# Create Conda environment
conda create --name AVSRCocktail python=3.11
conda activate AVSRCocktail
# Install FFmpeg, if it's not already installed.
conda install ffmpeg
# Install dependencies
pip install -r requirements.txt
```
## <a id="evaluation">2. Evaluation</a>
The evaluation script `script/evaluation.py` provides comprehensive evaluation capabilities for the AVSR Cocktail model on multiple datasets with various noise conditions and interference scenarios.
### Quick Start
**Basic evaluation on LRS2 test set:**
```sh
python script/evaluation.py --model_type avsr_cocktail --dataset_name lrs2 --set_id test
```
**Evaluation on AVCocktail dataset:**
```sh
python script/evaluation.py --model_type avsr_cocktail --dataset_name AVCocktail --set_id video_0
```
### Supported Datasets
#### 1. LRS2 Dataset
Evaluate on the LRS2 dataset with various noise conditions:
**Available test sets:**
- `test`: Clean test set
- `test_snr_n5_interferer_1`: SNR -5dB with 1 interferer
- `test_snr_n5_interferer_2`: SNR -5dB with 2 interferers
- `test_snr_0_interferer_1`: SNR 0dB with 1 interferer
- `test_snr_0_interferer_2`: SNR 0dB with 2 interferers
- `test_snr_5_interferer_1`: SNR 5dB with 1 interferer
- `test_snr_5_interferer_2`: SNR 5dB with 2 interferers
- `test_snr_10_interferer_1`: SNR 10dB with 1 interferer
- `test_snr_10_interferer_2`: SNR 10dB with 2 interferers
- `*`: Evaluate on all test sets and report average WER
**Example:**
```sh
# Evaluate on clean test set
python script/evaluation.py --model_type avsr_cocktail --dataset_name lrs2 --set_id test
# Evaluate on noisy conditions
python script/evaluation.py --model_type avsr_cocktail --dataset_name lrs2 --set_id test_snr_0_interferer_1
# Evaluate on all conditions
python script/evaluation.py --model_type avsr_cocktail --dataset_name lrs2 --set_id "*"
```
#### 2. AVCocktail Dataset
Evaluate on the AVCocktail cocktail party dataset:
**Available video sets:**
- `video_0` to `video_50`: Individual video sessions
- `*`: Evaluate on all video sessions and report average WER
The evaluation reports WER for three different chunking strategies:
- `asd_chunk`: Chunks based on Active Speaker Detection
- `fixed_chunk`: Fixed-duration chunks
- `gold_chunk`: Ground truth optimal chunks
**Example:**
```sh
# Evaluate on specific video
python script/evaluation.py --model_type avsr_cocktail --dataset_name AVCocktail --set_id video_0
# Evaluate on all videos
python script/evaluation.py --model_type avsr_cocktail --dataset_name AVCocktail --set_id "*"
```
### Configuration Options
#### Model Configuration
- `--model_type`: Model architecture to use (use `avsr_cocktail` for the AVSR Cocktail model)
- `--checkpoint_path`: Path to custom model checkpoint (default: uses pretrained `nguyenvulebinh/AVSRCocktail`)
- `--cache_dir`: Directory to cache downloaded models (default: `./model-bin`)
#### Processing Parameters
- `--max_length`: Maximum length of video segments in seconds (default: 15)
- `--beam_size`: Beam size for beam search decoding (default: 3)
#### Dataset Parameters
- `--dataset_name`: Dataset to evaluate on (`lrs2` or `AVCocktail`)
- `--set_id`: Specific subset to evaluate (see dataset-specific options above)
#### Output Options
- `--verbose`: Enable verbose output during processing
- `--output_dir_name`: Name of output directory for session processing (default: `output`)
### Advanced Usage
**Custom model checkpoint:**
```sh
python script/evaluation.py \
--model_type avsr_cocktail \
--dataset_name lrs2 \
--set_id test \
--checkpoint_path ./model-bin/my_custom_model \
--cache_dir ./custom_cache
```
**Optimized inference settings:**
```sh
python script/evaluation.py \
--model_type avsr_cocktail \
--dataset_name AVCocktail \
--set_id "*" \
--max_length 10 \
--beam_size 5 \
--verbose
```
### Output Format
The evaluation script outputs Word Error Rate (WER) scores:
**LRS2 evaluation output:**
```
WER test: 0.1234
```
**AVCocktail evaluation output:**
```
WER video_0 asd_chunk: 0.1234
WER video_0 fixed_chunk: 0.1456
WER video_0 gold_chunk: 0.1123
```
When using `--set_id "*"`, the script reports both individual and average WER scores across all test conditions.
## <a id="training">3. Training</a>
### Model Architecture
- **Encoder**: Pre-trained AV-HuBERT large model (`nguyenvulebinh/avhubert_encoder_large_noise_pt_noise_ft_433h`)
- **Decoder**: Transformer decoder with CTC/Attention joint training
- **Tokenization**: SentencePiece unigram tokenizer with 5000 vocabulary units
- **Input**: Video frames are cropped to the mouth region of interest using a 96 × 96 bounding box, while the audio is sampled at a 16 kHz rate
### Training Data
The model is trained on multiple large-scale datasets that have been preprocessed and are ready for the training pipeline. All datasets are hosted on Hugging Face at [nguyenvulebinh/AVYT](https://huggingface.co/datasets/nguyenvulebinh/AVYT) and include:
| Dataset | Size |
|---------|------|
| **LRS2** | ~145k samples |
| **VoxCeleb2** | ~540k samples |
| **AVYT** | ~717k samples |
| **AVYT-mix** | ~483k samples |
The information about these datasets can be found in the [Cocktail-Party Audio-Visual Speech Recognition](https://arxiv.org/abs/2506.02178) paper.
**Dataset Features:**
- **Preprocessed**: All audio-visual data is pre-processed and ready for direct input to the training pipeline
- **Multi-modal**: Each sample contains synchronized audio and video (mouth crop) data
- **Labeled**: Text transcriptions for supervised learning
The training pipeline automatically handles dataset loading and loads data in [streaming mode](https://huggingface.co/docs/datasets/stream). However, to make training faster and more stable, it's recommended to download all datasets before running the training pipeline. The storage needed to save all datasets is approximately 1.46 TB.
### Training Process
The training script is available at `script/train.py`.
**Multi-GPU Distributed Training:**
```sh
# Set environment variables for distributed training
export NCCL_DEBUG=WARN
export OMP_NUM_THREADS=1
export CUDA_VISIBLE_DEVICES=0,1,2,3
# Run with torchrun for multi-GPU training (using default parameters)
torchrun --nproc_per_node 4 script/train.py
# Run with custom parameters
torchrun --nproc_per_node 4 script/train.py \
--streaming_dataset \
--batch_size 6 \
--max_steps 400000 \
--gradient_accumulation_steps 2 \
--save_steps 2000 \
--eval_steps 2000 \
--learning_rate 1e-4 \
--warmup_steps 4000 \
--checkpoint_name avsr_avhubert_ctcattn \
--model_name_or_path ./model-bin/avsr_cocktail \
--output_dir ./model-bin
```
**Model Output:**
The trained model will be saved by default in `model-bin/{checkpoint_name}/` (default: `model-bin/avsr_avhubert_ctcattn/`).
#### Configuration Options
You can customize training parameters using command line arguments:
**Dataset Options:**
- `--streaming_dataset`: Use streaming mode for datasets (default: False)
**Training Parameters:**
- `--batch_size`: Batch size per device (default: 6)
- `--max_steps`: Total training steps (default: 400000)
- `--learning_rate`: Initial learning rate (default: 1e-4)
- `--warmup_steps`: Learning rate warmup steps (default: 4000)
- `--gradient_accumulation_steps`: Gradient accumulation (default: 2)
**Checkpoint and Logging:**
- `--save_steps`: Checkpoint saving frequency (default: 2000)
- `--eval_steps`: Evaluation frequency (default: 2000)
- `--log_interval`: Logging frequency (default: 25)
- `--checkpoint_name`: Name for the checkpoint directory (default: "avsr_avhubert_ctcattn")
- `--resume_from_checkpoint`: Resume training from last checkpoint (default: False)
**Model and Output:**
- `--model_name_or_path`: Path to pretrained model (default: "./model-bin/avsr_cocktail")
- `--output_dir`: Output directory for checkpoints (default: "./model-bin")
- `--report_to`: Logging backend, "wandb" or "none" (default: "none")
**Hardware Requirements:**
- **GPU Memory**: The default training configuration is designed to fit within **24GB GPU memory**
- **Training Time**: With 2x NVIDIA Titan RTX 24GB GPUs, training takes approximately **56 hours per epoch**
- **Convergence**: **200,000 steps** (total batch size 24) is typically sufficient for model convergence
## Acknowledgement
This repository is built using the [auto_avsr](https://github.com/mpc001/auto_avsr), [espnet](https://github.com/espnet/espnet), and [avhubert](https://github.com/facebookresearch/av_hubert) repositories.
## Contact
[email protected]