<!-- omit in toc -->

EAT: Self-Supervised Pre-Training with Efficient Audio Transformer

Guides

• Requirements and Installation
• Model Checkpoints
• Feature Extraction
• Data Preparation
• Pre-Training
• Fine-Tuning
• Inference and Evaluation

<!-- omit in toc -->

News 🔥

• [Update May. 3, 2025] 🎉🎉🎉 EAT now supports Hugging Face integration! You can extract features or run inference without relying on Fairseq — try EAT as your new audio encoder today!
• We release EAT-large (20 epochs) with SOTA performance on AS-2M, AS-20K, ESC-50 and SPC-2.
• Checkpoints and code are updated — EAT now seamlessly supports variable-length audio across training, extraction, inference, and evaluation.

<!-- omit in toc -->

Introduction

EAT is an audio SSL model with high effectiveness and efficiency during self-supervised pre-training. You can find details in the paper EAT: Self-Supervised Pre-Training with Efficient Audio Transformer.

Requirements and Installation

The minimum environment requirements are Python >= 3.8 and PyTorch >= 1.13. Please use pip < 24.1 due to dependency issues.

We now support Hugging Face integration — if you're only performing feature extraction or inference, you no longer need to install Fairseq!

🟡 For feature extraction or inference only (Hugging Face)

No Fairseq needed. Simply run:

git clone https://github.com/cwx-worst-one/EAT
cd EAT
pip install -r requirements.txt

🔵 For pre-training or fine-tuning (Fairseq-based)

You need to install Fairseq manually:

git clone https://github.com/pytorch/fairseq
cd fairseq
pip install --editable ./
git clone https://github.com/cwx-worst-one/EAT
pip install -r EAT/requirements.txt

Model Checkpoints

We provide several checkpoints for download, including both the original paper version and enhanced variants.

🔹 EAT-base (introduced in paper; for efficient pre-training)

• Pre-trained (10 epochs)
• Fine-tuned on AS-2M
• Fine-tuned on AS-20K

🔹 Updated & Recommended Versions

These enhanced versions feature extended pre-training or larger backbones.

• Checkpoints via Google Drive are compatible with Fairseq for further pre-training or fine-tuning.
• Hugging Face versions support direct usage via AutoModel.from_pretrained for feature extraction or inference.

| Version| 📦 Google Drive| 🤗 Hugging Face| |------------| :------------: | :--------: | | EAT-base (Epoch 30, Pre-trained)| Link | Link | | EAT-base (Epoch 30, Fine-tuned on AS-2M)| Link | Link | | EAT-large (Epoch 20, Pre-trained)| Link | Link | | EAT-large (Epoch 20, Fine-tuned on AS-2M) | Link | Link |

🧠 Browse collection on Hugging Face
⚠️ Note: Due to our limited AudioSet subset compared to other models, we recommend pre-training EAT on your own data for better performance.

📈 Performance Summary

| Model | Backbone | Params | Pre-train Epochs | AS-20K mAP (%) | AS-2M mAP (%) | |:-:|:-:|:-:|:-:|:-:|:-:| |EAT-base|ViT-B|88M|10|40.3 | 48.6| |EAT-base|ViT-B|88M|30|41.3 | 48.9| |EAT-large|ViT-L|309M|20|42.0 | 49.5|

Feature Extraction

We provide a simple script to extract audio features from the last layer of the EAT encoder. You can run feature extraction using either a Fairseq checkpoint or a Hugging Face model.

To get started, simply run:

bash EAT/scripts/feature_extract.sh 

For more detailed instructions, see the feature extraction guide.

Data Preparation

The main dataset in our experiment is AudioSet. Regrettably, we are unable to distribute the original audio files due to copyright restrictions.

However, you can access our data manifest here, which provides metadata and paths necessary for processing. We follow the file format in wav2vec and data2vec, where .tsv format file is for index while .lbl and .csv format files are specific for classification task. You are free to modify the files to suit your own datasets or experimental needs.

Pre-Training

Our codes are adapted from Audio-MAE and data2vec. The default configuration file for pre-training is pretraining_AS2M.yaml. To pre-train the EAT model on AudioSet, simply run the following script:

bash EAT/scripts/pretraining_AS2M.sh 

If you wish to pre-train the model on a different dataset where audio clips are not fixed to 10 seconds, please refer to the feature extraction guide for detailed instructions on how to adjust the target length accordingly.

Fine-Tuning

We use finetuning.yaml as the default configuration file for fine-tuning. To fine-tune the EAT model in different downstream tasks, you could run the script finetuning_{task}.sh, where {task} refers to one of the supported datasets, including AS20K, AS2M, ESC50, and SPCv2. For example, you can fine-tune EAT on AS20K by executing:

bash EAT/scripts/finetuning_AS20K.sh

Inference and Evaluation

We support both local inference and loading models via Hugging Face. To run inference on a single AudioSet audio clip using fine-tuned EAT models, you may use our checkpoint fine-tuned on AS-2M (recommended) or AS-20K. Alternatively, you can load the models directly from Hugging Face.

To start inference, run:

bash EAT/scripts/inference.sh 

An example output is as follows:

# top_k_prediction = 12
************ Acoustic Event Inference ************
LABEL                          PREDICTION
Percussion                     0.523
Drum kit                       0.437
Vibraphone                     0.420
Drum                           0.316
Music                          0.303
Snare drum                     0.277
Glockenspiel                   0.225
Marimba, xylophone             0.223
Cymbal                         0.213
Bass drum                      0.207
Hi-hat                         0.196
Mallet percussion              0.170
**************************************************

To evaluate on the full AudioSet evaluation set, use:

bash EAT/scripts/eval.sh 

This will report the evaluation value of mAP on the test set. Per-class AP scores are saved to ./EAT/ap_log.txt. You could also refer to our results of finetuned EAT models on evaluation set of Audioset in ./EAT/results.

<!-- omit in toc -->

Performance

Pre-training on AS-2M, EAT gains state-of-the-art (SOTA) performance on several audio and speech classification datasets including AS-20K, AS-2M, ESC-50 and SPC-2.

<!-- omit in toc -->

Efficiency

EAT achieves a total pre-training time reduction of ~15x compared to BEATs and ~10x relative to Audio-MAE. The full pre-training on AS-2M requires only 10 epochs.

<!-- omit in toc -->

Experiment Logs

We report the experiment logs using wandb. We have published a short WandB report detailing the training process and performance metrics of the EAT model. You could visit it here.

<!-- omit in toc -->

TODO

• [x] Release the final EAT large
• [x] Update codes and checkpoints for friendly usage
• [x] Provide model support on Hugging Face
• [ ] Release the Docker image

Acknowledgement

Our codebase is based on the awesome [Audio-MAE](https://github.com/facebookres