Comparison of Machine Learning Classifier Models for Microphone Array based Acoustic Source Localisation

This work addresses a long-standing challenge by employing machine learning to locate an acoustic source by directly estimating its location based on recorded data from a linear array of microphones spaced 5 cm apart. Various machine learning classifier methods are employed to solve the supervised learning problem of acoustic source localisation. In this article, five cutting-edge machine learning classification techniques are examined: XGboost, adaboost, gradient boosting, decision tree (DT), and random forest (RF) —utilized to locate sound sources using experimental data. To improve the classification of the angular position of the sound source in an interior environment, the five machine learning models’ performances are compared. A linear array of sixteen microphones has been used in experiments to collect wide band audio signals from an acoustic source positioned at different angles. Single input samples from each sensor are utilized simultaneously to train machine learning models that are meant to locate the acoustic source. Ultimately, an independent test dataset were used to assess the models’ performance. For the comparative analysis, the accuracy of all the machine learning models are taken into consideration. Additionally, the receiver operating characteristic curve was used to evaluate these models. The findings indicated that the XGboost classifier produced better outcomes with a higher accuracy. Adaboost was the least successful machine learning model for sensor array based acoustic localisation, underperforming in every pitch.