Need help with seld-net?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

sharathadavanne
194 Stars 49 Forks Other 25 Commits 7 Opened issues

Description

Sound event localization, detection, and tracking of multiple overlapping and moving sources in 2D spherical space using convolutional recurrent neural network

Services available

!
?

Need anything else?

Contributors list

No Data

Sound event localization, detection, and tracking of multiple overlapping stationary and moving sources using convolutional recurrent neural network (SELDnet)

We have formalized this work as a research challenges at IEEE AASP workshop DCASE. 1. In the first version of the sound event localization and detection (SELD) challenge at DCASE 2019. We provided dataset with stationary sources in multiple reverberant scenarios. 2. In the second version of the sound event localization and detection (SELD) challenge at DCASE 2020. We provided dataset with both stationary and moving sources in multiple reverberant scenarios.

Sound event localization, detection and tracking (SELDT) is the combined task of identifying the temporal onset and offset of a sound event, tracking the spatial location when active, and further associating a textual label describing the sound event. We first presented the SELDnet for static scenes with spatially stationary sources in IEEExplore (also available on Arxiv). Thereafter, we presented the performance of SELDnet on dynamic scenes with sources moving with different angular velocities here. We observed that the recurrent layers are crucial for tracking of sources, and perform comparable tracking as bayesian trackers such as RBMCDA particle filter (Code available here). We are releasing a simple vanila code without much frills and the related datasets here. If you want to read more about the general literature of SELDT, you can refer here.

The work presented in the above two papers is an extension of the previous multichannel sound event detection, and direction of arrival estimation papers listed below.

  1. Sound event detection (SED)

  2. Direction of arrival (DOA) estimation

If you are using this code or the datasets in any format, then please consider citing the following papers

Sharath Adavanne, Archontis Politis, Joonas Nikunen, and Tuomas Virtanen, "Sound event localization and detection of overlapping sources using convolutional recurrent neural network" in IEEE Journal of Selected Topics in Signal Processing (JSTSP 2018)

Sharath Adavanne, Archontis Politis and Tuomas Virtanen, "Localization, detection, and tracking of multiple moving sources using convolutional recurrent neural network" submitted in IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA 2019)

More about SELDnet

The proposed SELDnet architecture is as shown below. The input is the multichannel audio, from which the phase and magnitude components are extracted and used as separate features. The proposed method takes a sequence of consecutive spectrogram frames as input and predicts all the sound event classes active for each of the input frame along with their respective spatial location, producing the temporal activity and DOA trajectory for each sound event class. In particular, a convolutional recurrent neural network (CRNN) is used to map the frame sequence to the two outputs in parallel. At the first output, SED is performed as a multi-label multi-class classification task, allowing the network to simultaneously estimate the presence of multiple sound events for each frame. At the second output, DOA estimates in the continuous 3D space are obtained as a multi-output regression task, where each sound event class is associated with three regressors that estimate the 3D Cartesian coordinates x, y and z of the DOA on a unit sphere around the microphone.

The SED output of the network is in the continuous range of [0 1] for each sound event in the dataset, and this value is thresholded to obtain a binary decision for the respective sound event activity as shown in figure below. Finally, the respective DOA estimates for these active sound event classes provide their spatial locations.

The figure below visualizes the SELDnet input and outputs for simulated datasets with maximum one (O1) and two (O2) temporally overlapping and stationary sound events. The horizontal-axis of all sub-plots for a given dataset represents the same time frames, the vertical-axis for spectrogram sub-plot represents the frequency bins, vertical-axis for SED reference and prediction sub-plots represents the unique sound event class identifier, and for the DOA reference and prediction sub-plots, it represents the distance from the origin along the respective axes. The 'o' markers in left figure and '•' markers in right figure visualize both the groundtruth labels and predictions of DOA and SED for O1 and O2 datasets. The − markers in the left figure visualizes the results for test data with unseen DOA labels (shifted by 5 degree along azimuth and elevation). The figures represents each sound event class and its associated DOA outputs with a unique color.

Similarly, the figure below visualizes the SELDnet input and outputs for moving source dataset with maximum two temporally overlapping sound events.

DATASETS

We are releasing all the simulated datasets and the small real-life dataset without ambiance used in the paper on zenodo.org. These datasets are in the range of 30-45 GB and fit within the dataset budget of zenodo.org. The larger datasets can be shared upon request. The first five datasets consist of stationary point sources each associated with a spatial coordinate. Whereas the last two datasets consists of moving point sources with varying angular velocities.

The datasets released are 1. ANSIM (TUT Sound Events 2018 - Ambisonic, Anechoic and Synthetic Impulse Response Dataset) DOI 2. RESIM (TUT Sound Events 2018 - Ambisonic, Reverberant and Synthetic Impulse Response Dataset) DOI 3. CANSIM (TUT Sound Events 2018 - Circular array, Anechoic and Synthetic Impulse Response Dataset) DOI 4. CRESIM (TUT Sound Events 2018 - Circular array, Reverberant and Synthetic Impulse Response Dataset) DOI 5. REAL (TUT Sound Events 2018 - Ambisonic, Reverberant and Real-life Impulse Response Dataset) DOI - Real-life impulse responses to simulate custom SELD datasets DOI 6. MANSIM (TAU Moving Sound Events 2019 - Ambisonic, Anechoic, Synthetic Impulse Response and Moving Sources Dataset) DOI 7. MREAL (TAU Moving Sound Events 2019 - Ambisonic, Reverberant, Real-life Impulse Response and Moving Sources Dataset) DOI

All the datasets contain three sub-datasets with maximum one (ov1), two (ov2) and three (ov3) temporally overlapping sound events. Each of these sub-datasets have three cross-validation splits (split1, split2 and split3). In total each dataset has nine splits saved as separate zip files. In order to test the SELDnet code you don't have to download the entire dataset. You can simply download one of the zip files and train the SELDnet for the respective overlap (ov) and split (split).

Baseling tracking method

In order to compare the tracking performance of SELDnet, we used the parameteric method comprising of MUSIC for frame-wise DOA estimation and particle filter with Rao-Blackwellized Monte Carlo Data Association (RBMCDA). This RBMCDA particle filter has also been made publicly available here.

Getting Started

This repository consists of multiple Python scripts forming one big architecture used to train the SELDnet. * The batchfeatureextraction.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 clsfeatureclass.py script has routines for labels creation, features extraction and normalization. * The clsdatagenerator.py script provides feature + label data in generator mode for training. * The kerasmodel.py script implements the SELDnet architecture. * The evaluationmetrics.py script, implements the core metrics from sound event detection evaluation module http://tut-arg.github.io/sed_eval/ and the DOA metrics explained in the paper * The seld.py is a wrapper script that trains the SELDnet. The training stops when the SELD error (check paper) stops improving. * The utils.py script has some utility functions.

If you are only interested in the SELDnet model then just check the keras_model.py script.

Prerequisites

The requirements.txt file consists of the libraries and their versions used. The Python script is written and tested in 3.7.3 version. You can install the requirements by running the following line

pip install -r requirements.txt

Training the SELDnet on the datasets

The SELDnet code trains the network for a given dataset (ansim, resim, cansim, cresim or real), overlap (ov1, ov2 or ov3) and split (split1, split2 or split3) at a time. In order to quickly train SELDnet follow the steps below.

  • For the chosen dataset (ansim or resim or ..), overlap (1, 2 or 3) and split (1, 2 or 3), download the respective zip file. This contains both the audio files and the respective metadata. Unzip the files under the same 'basefolder/', ie, if you are downloading overlap 1 and split 1 of the ansim dataset, then the 'basefolder/' should have two folders - 'wavov1split130db/' and 'descov1_split1/' after unzipping.

  • Now update the respective dataset path in clsfeatureclass.py script. For the above example, you will change line 22 to

    self._base_folder = 'base_folder/'
    (below
    dataset == 'ansim'
    ). The normalized features, and labels are written in the same folder, so make sure you have sufficient space for it.
  • Extract features from the downloaded dataset by running the batchfeatureextraction.py script. First, update the datasetname, overlap and split values based on the downloaded dataset. See the python file for more comments. You can now run the script as shown below. This will create a new folder inside the 'basefolder/' and dump the normalized features here. Since feature extraction is a one-time thing, this script is standalone and does not use the parameter.py file. As you can see in the script, you can extract features for all possible overlaps and splits in one shot with this script.

python batch_feature_extraction.py
  • Update the parameters

    dataset
    ,
    overlap
    and
    split
    in parameter.py script based on the downloaded dataset. You can now train the SELDnet using default parameters using
    python seld.py
    
  • Additionally, you can add/change parameters by using a unique identifier <task-id> in if-else loop as seen in the parameter.py script and call them as following

    python seld.py  
    
    Where <job-id> is a unique identifier which is used for output filenames (models, training plots). You can use any number or string for this.
  • By default, the code runs in

    quick_test = True
    mode. This trains the network for 2 epochs on only 2 mini-batches. Once you get to run the code sucessfully, set
    quick_test = False
    in parameter.py script and train on the entire data.

License

This repository is licensed under the TUT License - see the LICENSE file for details

Acknowledgments

The research leading to these results has received funding from the European Research Council under the European Unions H2020 Framework Programme through ERC Grant Agreement 637422 EVERYSOUND.

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.