Sound Event Localization and Detection using Multi-Target TDOA Features

This repository implements the audio-only part of our DCASE 2024 challenge submission.

@techreport{Berg_LU_task3_report,
    Author = "Berg, Axel and Engman, Johanna and Gulin, Jens and Astrom, Karl and Oskarsson, Magnus",
    title = "THE LU SYSTEM FOR DCASE 2024 SOUND EVENT LOCALIZATION AND DETECTION CHALLENGE",
    institution = "DCASE2024 Challenge",
    year = "2024",
}

A more detailed description of the method can be found in the arXiv paper.

@article{berg2024learning,
  title={Learning Multi-Target TDOA Features for Sound Event Localization and Detection},
  author={Berg, Axel and Engman, Johanna and Gulin, Jens and {\AA}str{\"o}m, Karl and Oskarsson, Magnus},
  journal={arXiv preprint arXiv:2408.17166},
  year={2024}
}

This repository is based on the DCASE 2024 SELD challenge baseline implementation.

Please visit the official webpage of the DCASE 2024 Challenge for details missing in this repo.

Datasets

Download the MIC dataset from here: Sony-TAu Realistic Spatial Soundscapes 2023 (STARSS23)

In order to train the models with (optional) additional simulated recordings, then download the official synthetic dataset from here: [DCASE2024 Task 3] Synthetic SELD mixtures for baseline training

In our challenge submission, we also generated extra synthetic data. For this, we refer to Spatial Scaper

In addition, we use channel swapping for data augmentation. We used SoundQ in order to apply this to the real recordings.

Getting Started

This repository consists of multiple Python scripts.

• The batch_feature_extraction.py is a standalone wrapper script, that extracts the features, labels, and normalizes the training and test split features for a given dataset. Make sure you update the location of the downloaded datasets before.
• The parameter.py script consists of all the training, model, and feature parameters. If a user has to change some parameters, they have to create a sub-task with unique id here. Check code for examples.
• The cls_feature_class.py script has routines for labels creation, features extraction and normalization.
• The cls_vid_features.py script extracts video features for the audio-visual task from a pretrained ResNet model.
• The cls_data_generator.py script provides feature + label data in generator mode for training.
• The seldnet_model.py script implements the SELDnet baseline architecture.
• The CST-Former architecture is found in cst_former/
• The NGCC-PHAT architecture is found in ngcc/
• The SELD_evaluation_metrics.py script implements the metrics for joint evaluation of detection and localization.
• The train_seldnet.py is a wrapper script that performs pre-training of NGCC-PHAT and end-to-end training of SELDNET/CST-Former
• The cls_compute_seld_results.py script computes the metrics results on your DCASE output format files.

Feature extraction

• For the MIC dataset, download the zip files. This contains both the audio files and the respective metadata.
• Unzip the files under data_2024/
• Extract features from the downloaded dataset by running the batch_feature_extraction.py script. Run the script as shown below. This will dump the normalized features and labels in the feat_label_dir folder. The Python script allows you to compute all possible features and labels. You can control this by editing the parameter.py script before running the batch_feature_extraction.py script.

GCC + MS:

python batch_feature_extraction.py 6

SALSA-LITE:

python batch_feature_extraction.py 7

NGCC + MS:

python batch_feature_extraction.py 9

NGCC-PHAT computes features from raw audio on the fly. However, using params['saved_chunks'] = True, will convert the .wav.files to .npy, which makes them faster to load.

NGCC Pre-Training

• We provide a pre-trained NGCC-PHAT feature extractor in models/9_tdoa-3tracks-16channels.h5. If you want to perform ADPIT pre-training of the NGCC-PHAT network yourself, run

python train_seldnet.py 9 my_pretraining

SELD Training

• Finally, you can now train the SELDnet Baseline or CST-Former using.

python3 train_seldnet.py <task-id> <job-id> <seed>

The default training setups are

Baseline w/ NGCC+MS

python3 train_seldnet.py 10 my_experiment

Baseline w/ GCC+MS

python3 train_seldnet.py 6 my_experiment

Baseline w/ SALSA-Lite

python3 train_seldnet.py 7 my_experiment

CST-Former w/ NGCC+MS

python3 train_seldnet.py 333 my_experiment

CST-Former w/ GCC+MS

python3 train_seldnet.py 33 my_experiment

CST-Former w/ SALSA-Lite

python3 train_seldnet.py 34 my_experiment

Results on development dataset

The evaluation metric scores for the test split of the development dataset with the provided pre-trained models are given below.

Model	F20°	DOAECD	RDECD	weights
CST-Former Small w/ NGCC	28.3%	23.7°	0.42	models/333_cst-3t16c.h5
CST-Former Large w/ NGCC	31.3%	21.6°	0.48	models/333_cst-3t16c-large.h5

Acknowledgement

This repository is based on the DCASE 2024 SELD challenge baseline implementation.

The original NGCC implementation can be found here.

The CST-Former implementation is borrowed from this repository. Please cite their work as well:

@INPROCEEDINGS{shul2024cstformer,
  author={Shul, Yusun and Choi, Jung-Woo},
  booktitle={ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={CST-Former: Transformer with Channel-Spectro-Temporal Attention for Sound Event Localization and Detection}, 
  year={2024},
  pages={8686-8690},
  doi={10.1109/ICASSP48485.2024.10447181}}

License

This repo and its contents have the MIT License.