This project provides functionality for generating synthetic room acoustic datasets, which can be used for training and evaluating machine learning models. Furthermore, we take interest in asynchronous microphone networks, which may be also sometimes referred to as Wireless Acoustic Sensor Networks (WASNs). Every dataset sample is generated in two steps, firstly by simulating the acoustic propagation from the sound sources to all the microphones present in the room, and later by simulating the network propagation and microphone characteristics through the use of fractional delay filtering.
So far, Sydra is only able to generate datasets containing multiple microphones and sources. The parameters of these microphones may be either fixed or random, as well as the room dimensions and reverberation time (rt60). The input source signal may either be noise, or a directory containing signal samples (such as speech) may be provided.
We provide a SydraDataset class, a Pytorch Dataset, which allows you to easily load and train models using Pytorch.
We recommend creating a virutal environment for this project. If you are using Anaconda, this can be achieved by running the following command:
conda create -n sydra python=3.8
Then, activate the environment:
conda activate sydra
Then, install the requirements:
pip install -r requirements.txt
Finally, clone the package:
git clone https://github.com/SOUNDS-RESEARCH/sydra --recurse-submodules
The --recurse-submodules flag is necessary to clone the Pyroomasync submodule, a wrapper on top of Pyroomacoustics which facilitates generating behaviour
common to WASNs (Microphones with different sampling rates, gains, delay).
You can configure your simulation in the config/config.yaml file. Most parameters in the config file have an inline comment describing it. Then, you can create a dataset by running:
python main.py dataset_dir=/path/to/dataset n_samples=1000,
where dataset_dir is the directory where the dataset will be saved, and n_samples is the number of samples to generate.
Every SydraDataset contains a metadata.[csv|json|yaml] file in its root, which contains the annotations for every dataset sample.
A SydraDataset also contains a samples/ directory, which in turn contains many directories. Every one of such directories contains the recordings referring to each microphone.
- Eric Grinstein
- Derin Ak
- Prof. Patrick A. Naylor
Imperial College London