Sound Reconstruction

This repository shows the data and demo code:

• Park JY, Tsukamoto M, Tanaka M, Kamitani Y (2025) Natural sounds can be reconstructed from human neuroimaging data using deep neural network representation. PLOS Biology 23(7): e3003293. https://doi.org/10.1371/journal.pbio.3003293
• Park, JY, Tsukamoto M, Tanaka M, & Kamitani Y (2023) Sound reconstruction from human brain activity via a generative model with brain-like auditory features (arXiv:2306.11629). arXiv. http://arxiv.org/abs/2306.11629

Dataset

• Raw fMRI data: OpenNeuro
• Preprocessed fMRI data, DNN features extracted from sound clips: figshare
• Trained transformer models: figshare
• Stimulus sound clips: Refer to data/README.md .

Code

Setup

1. Clone this SoundReconstruction repository to your local machine (GPU machine preferred).

git clone [email protected]:KamitaniLab/SoundReconstruction.git

2. Create conda environment using the specvqgan.yaml.

conda env create --name specvqgan -f specvqgan.yaml 
python -c "import torch; print(torch.cuda.is_available())"
# True

3. Clone SpecVQGAN repository next to SoundReconstruction directory. Please use the following fork repository instead of the original SpecVQGAN repository because the path of the Transformer configuration file has been rewritten.

git clone [email protected]:KamitaniLab/SpecVQGAN.git

Download datasets and models

See data/README.md.

Usage

We provide scripts that reproduce main results in the original paper. Please execute the sh files in the following order.

1. Train feature decoders to predict the VGGishish features.

./1_train_batch.sh

2. Using the decoders trained in step.1, perform feature predictions. (Perform the prediction for the attention task dataset at the same time.)

./2_test_batch.sh

3. Validate the prediction accuracy of predicted features.

./3_eval_batch.sh

Visualize the prediction accuracy with the following notebook. This notebook draws Fig.3D and Fig.3E of the original paper.

feature_decoding/makefigures_featdec_eval.ipynb

4. Reconstruct sound clips using predicted features.

./4_recon_batch.sh

5. Validate the quality of reconstructed sound.

./5_recon_eval_batch.sh 

Visualize the reconstruction quality with the following notebooks. These notebooks draws Fig.4C and Fig.8C of the original paper.

reconstruction/makefigures_recon_eval.ipynb
reconstruction/makefigures_recon_eval_attention.ipynb