EVALUATING PYROOMACOUSTICS BEAMFORMING AND DIRECTION-OF-ARRIVAL IMPLEMENTATIONS ON THE AIRA CORPUS
Budget: $10 – $30 USD
I In this project from the speech signal processing and microphone array processing domain, we want to evaluate the "pyroomacoustics" package on the AIRA (Acoustic interactions for robot audition) corpus.
The purpose of this project is to compare prominent implementations of microphone array processing algorithms and to evaluate them on an established corpus.
Therefore, this project serves as a theoretical overview of the current state-of-the-art in microphone array processing with a practical example.
We will focus on pyroomacoustics' beamforming and direction-of-arrival algorithms.
The implementation will be done in jupyter notebook in a step-by-step fashion with sufficient explanations in between.
We base the structure of the implementation on the official pyroomacoustics demo (see links section).
Following the documentation of the AIRA corpus, we can define the room configurations in a JSON file.
There is already an example in "rooms.json" that resembles the "anechoic chamber" example from the AIRA corpus.
We still need to add the Cafeteria, Department Store, Hall, Office A, Office B, and Office C configurations.
Where the documentation is incomplete, use estimated distances based on the graphics provided.
In the attached notebook, we have already provided a code snippet to process room data and plot the gold standard configurations from AIRA.
We loop over the room data to process the different microphone configurations, for which we then add sound sources.
Once we have configured the sources as shown in the code example, we want to evaluate different algorithms and compare whether the algorithms can reproduce the sound source locations / DOA as outlined in the AIRA documentation.
For that, we use the beamforming and DOA algorithms to estimate the locations of sound sources based on the audio recordings provided by AIRA.
What follows is a comparison of the results and a discussion about the room definition procedure and AIRA's documentation.
We have intended the report to be structured as follows (suggestions and edits welcome):
x Abstract
x Introduction
x Theory and tools
- Beamforming algorithms
o Delay-and-sum
o MVDR
- Direction-of-arrival algorithms
o CSSM
o FRIDA
o MUSIC
o SRP-PHAT
o TOPS
o WAVES
- AIRA corpus
- pyroomacoustics package
x Methodology
- Room definitions
- Serial processing for rooms and configurations
- Comparing predictions of different algorithms with the original AIRA data
x Results
- Beamforming performance
- DOA performance
x Discussion
x Conclusion
x References
x Appendix
Links:
pyroomacoustics repository: https://github.com/LCAV/pyroomacoustics
pyroomacoustics demo: https://mybinder.org/v2/gh/LCAV/pyroomacoustics/master?filepath=notebooks%2Fpyroomacoustics_demo.ipynb
AIRA dataset: https://aira.iimas.unam.mx/
AIRA documentation: https://aira.iimas.unam.mx/aira.pdf
The purpose of this project is to compare prominent implementations of microphone array processing algorithms and to evaluate them on an established corpus.
Therefore, this project serves as a theoretical overview of the current state-of-the-art in microphone array processing with a practical example.
We will focus on pyroomacoustics' beamforming and direction-of-arrival algorithms.
The implementation will be done in jupyter notebook in a step-by-step fashion with sufficient explanations in between.
We base the structure of the implementation on the official pyroomacoustics demo (see links section).
Following the documentation of the AIRA corpus, we can define the room configurations in a JSON file.
There is already an example in "rooms.json" that resembles the "anechoic chamber" example from the AIRA corpus.
We still need to add the Cafeteria, Department Store, Hall, Office A, Office B, and Office C configurations.
Where the documentation is incomplete, use estimated distances based on the graphics provided.
In the attached notebook, we have already provided a code snippet to process room data and plot the gold standard configurations from AIRA.
We loop over the room data to process the different microphone configurations, for which we then add sound sources.
Once we have configured the sources as shown in the code example, we want to evaluate different algorithms and compare whether the algorithms can reproduce the sound source locations / DOA as outlined in the AIRA documentation.
For that, we use the beamforming and DOA algorithms to estimate the locations of sound sources based on the audio recordings provided by AIRA.
What follows is a comparison of the results and a discussion about the room definition procedure and AIRA's documentation.
We have intended the report to be structured as follows (suggestions and edits welcome):
x Abstract
x Introduction
x Theory and tools
- Beamforming algorithms
o Delay-and-sum
o MVDR
- Direction-of-arrival algorithms
o CSSM
o FRIDA
o MUSIC
o SRP-PHAT
o TOPS
o WAVES
- AIRA corpus
- pyroomacoustics package
x Methodology
- Room definitions
- Serial processing for rooms and configurations
- Comparing predictions of different algorithms with the original AIRA data
x Results
- Beamforming performance
- DOA performance
x Discussion
x Conclusion
x References
x Appendix
Links:
pyroomacoustics repository: https://github.com/LCAV/pyroomacoustics
pyroomacoustics demo: https://mybinder.org/v2/gh/LCAV/pyroomacoustics/master?filepath=notebooks%2Fpyroomacoustics_demo.ipynb
AIRA dataset: https://aira.iimas.unam.mx/
AIRA documentation: https://aira.iimas.unam.mx/aira.pdf