SoundLoc-EQ v1.0.0

[liincu4n]

SoundLoc-EQ v1.0.0

Release Date: March 14, 2025

We’re excited to announce the first official release of SoundLoc-EQ, a framework for unbiased sound source localization in circular spaces using deep learning. This release marks the completion of the project, delivering a robust solution for audio localization tasks.

Key Features

• Unbiased Encoding and Decoding: The core encoding_decoding.py module introduces unbiased_encoding and weighted_adjacent_decoding to eliminate quantization errors in 360° sound localization.
• Circular Space Support: Handles 360-degree localization with a grid resolution of 72 cells (5 degrees each).
• Distributed Training: Supports multi-GPU training with PyTorch Lightning’s DDP strategy.
• Comprehensive Evaluation: Tested on datasets like Libri_Circular_4 and Libri_adhoc_nodes10_splited, with robust performance in noisy and reverberant conditions.

What’s New

• Initial release of the complete SoundLoc-EQ framework.
• Full implementation of training (train.py) and evaluation (eval.py) pipelines.
• Detailed documentation in README.md, including setup, usage, and configuration instructions.
• Pre-trained model checkpoint (1hot_72.ckpt) included for immediate evaluation.

Installation

1. Clone the repository:

   git clone https://github.com/MetaAcoustix/SoundLoc-EQ.git
   cd SoundLoc-EQ

2. Install dependencies:

   pip install &&..

Follow the README.md for dataset setup and further instructions.

Usage

• Train: Run python train.py to train the SS2net model on your dataset.
• Evaluate: Run python eval.py to test the PhaseModel with the provided checkpoint.

Known Issues

• Dataset paths in train.py and eval.py are hardcoded (e.g., /home/liincuan/...). Update these to match your local environment.
• GPU support requires CUDA-compatible PyTorch; CPU fallback is available but slower.

Future Plans

• Add a Project Report file for easier dependency management.

Acknowledgments

• Thanks to the dataset providers for Libri_Circular_4 and Libri_adhoc_nodes10_splited.
• Inspired by research in unbiased label distribution learning.

For more details, check the README. We welcome feedback via GitHub issues!

---

This discussion was created from the release SoundLoc-EQ v1.0.0.