Ad hoc microphone array localizations using ML

Microphone array localization is the task of finding the relative positions of microphones in relation to each other.
It is important to perform spatial signal processing tasks, such as beamforming.
Microphone array that doesn’t have any specific structure and can’t be represented as a parametric equation are called ad-hoc arrays.
Ad-hoc arrays are very difficult to localize and calibrate since there are a lot of parameters that need to be estimated and they don’t corelate.
Automatic calibration without any prior knowledge is possible, given enough sound events in the scene.
It is done by solving an inverse problem that gets the positions.
In this work we explore a new framework to calibrate an ad-hoc array using machine learning.
Using unsupervised machine learning we take the microphone signals, feed them into a neural network to estimate the sought after parameters.
The estimated parameters are fed into a derivable sound propagation simulator to reconstruct the microphone signals.
Comparing the input signal and reconstructed signal allows us to calculate error and iterate to converge into a better estimation.
This allows a flexible progressive calibration process that takes into account all the information available in the microphone signal.

M.Sc. student under the supervision of Prof. Yoash Levron.