A package to localise sources in overlapping multi-channel audio data.
The name pydatemm refers to the DATEMM (Scheuing & Yang 2008) algorithm that was the original implementation.
The current version of this package implements the Compatibility-Conflict-Graph (CCG) (Kreißig & Yang 2013) algorithm.
The CCG and linear algebra localisation scripts are optimised for speed and written in C++ (python-C++ communication
through the cppyy package.
The Compatibility-Conflict-Graph algorithm by Kreißig & Yang is the successor of the DATEMM algorithm (Scheuing & Yang 2008), which builds zero-sum time-difference-of-arrival (TDOA) triplets, and 'fuses' compatible triplets together to make a bigger TDOA graph.
In CCG, zero-sum triplets are first built systematically by investigating the presence of zero-sum 'fundamental loops'. Then, based on the values and node identities, all zero-sum loops are compared with each other and a large N_loop x N_loop binary matrix is filled to indicate compatibility or a conflict. A compatibility indicates two triplets can be fused together, while a conflict indicates they can't. The CCG algorithm finds the largest subsets of unique compatible fundamental-loops, that thus result in >=4 channel TDOA graphs that can then be used to localise sources.
This package is rather under-documented from a user-perspective, and active development was stopped ~2023.
The heart of the package is the `generate_candidate_sources` function in the `source_generation` module.
- For examples scripts that use the
`generate_candidate_sources`see the`examples/`folder, and in specific: `ushichka_2018-08-17_speakerplayback.py`
- For examples of bash scripts (ending with
`.sh`) that were used to run the code on computing clusters with the Slurm job-manager check out: `1529543496_origxyz.sh`
Update 2025 January: This package is not very user-installable yet, and documentation may be patchy. This is a proof-of-principle implementation that didn't go too far when I realised the DATEMM/CCG class methods are rather slow when you need to start tracking multiple sources ( >=3 sources) with multiple microphones (>=12 mics). The code does work though, and the main issue is that there are a lot of false positives (and some false negatives too) - which makes cleaning the data hard unless you have another sensor modality to support the output (e.g. camera based trajectories to constrain audio source locations).
- Scheuing, J., & Yang, B. (2008). Disambiguation of TDOA estimation for multiple sources in reverberant environments. IEEE transactions on audio, speech, and language processing, 16(8), 1479-1489.
- Kreißig, M., & Yang, B. (2013, May). Fast and reliable TDOA assignment in multi-source reverberant environments. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing (pp. 355-359). IEEE.
- Free software: MIT license
- Documentation: https://pydatemm.readthedocs.io.
This package was created with Cookiecutter and the audreyr/cookiecutter-pypackage project template.