Inference Dataset

Subset of MeshRIR Dataset

Here we maintain a subset of the MeshRIR dataset, curated and maintained for reproducibility purposes. This dataset is a reduced version of the original MeshRIR dataset, specifically tailored to a particular source/receiver configuration and downsampled to a sampling rate of $fs = 8000$ Hz.

Dataset Overview:
The MeshRIR dataset is a collection of acoustic room impulse responses (RIRs) recorded at finely meshed grid points. It serves as a valuable resource for evaluating sound field analysis and synthesis methods. Our curated subset focuses on a specific source/receiver configuration, enabling researchers and practitioners to reproduce and verify results for this targeted scenario.

Contents:
The release includes the following:

Acoustic Room Impulse Responses (RIRs):

• A carefully selected subset of RIRs corresponding to a particular source/receiver configuration.
• The RIRs are downsampled to a uniform sampling rate of fs = 8000 Hz, reducing the dataset's overall size while maintaining essential acoustic characteristics.

Original MeshRIR Dataset:
The MeshRIR dataset, from which this subset is derived, is provided by Original Dataset Source. It consists of RIRs in a 3D cuboidal region from a single source and IRs in a 2D square region from an array of 32 sources.

Citation:
If you use this subset of the MeshRIR dataset in your research or publication, we kindly request that you cite the original MeshRIR dataset using the respective DOI reference:

• Original MeshRIR Dataset: DOI

Acknowledgments:
We express our gratitude to the creators of the original MeshRIR dataset for providing a valuable resource to the research community. This curated subset is an effort to contribute to the promotion of reproducibility and open science in acoustic research.

For any questions or issues related to this dataset, please feel free to contact us through [xenoka@dtu.dk].

Pre-trained GAN generator weights

Pre-trained GAN Generator Weights for Inference

This release availability of pre-trained weights for the Generative Adversarial Network (GAN) generator, as described in the publication titled "Generative adversarial networks with physical sound field priors" by Xenofon Karakonstantis and Efren Fernandez-Grande. These weights have been trained on a large dataset, and we are now providing them to the community for inference purposes.

GAN Generator Overview:
The GAN generator is a deep neural network architecture designed to generate synthetic data that closely resembles the distribution of real data. This pre-trained GAN generator has been trained to achieve diverse sound fields in accordance with the methods presented in the publication.

• The generator model is implemented in TensorFlow, making it compatible with a wide range of deep learning frameworks.

Usage Instructions:
You can easily utilise the pre-trained GAN generator for sound field inference. Below are the steps to get started:

1. Install the required dependencies, including TensorFlow [2.8].

2. Download the pre-trained generator weights and unzip file

3. Load the weights into your TensorFlow-based project using the provided code snippet:

   import tensorflow as tf
   
   # Define your GAN generator architecture (ensure it matches the architecture used during training)
   generator = YourGANGenerator()
   
   # Load pre-trained weights
   generator.load_weights("path/to/pretrained_weights")
   
   # Generate synthetic sound fields using the generator
   synthetic_soundfield = generator(noise_input)