Abstract
In-situ audio recordings for forensic analysis are generally made using purpose-installed microphones in order to capture speech and other relevant sounds in an environment. Often, post-processing efforts are focused on cleaning these recordings from interference. An acoustic scene is made up of audio objects and their not only contents, but also positions carry evidential importance. Therefore, in addition to cleaning target sources, detecting their locations may be required. This paper focuses on the performance of a closed-form acoustic scene decomposition technique, which blindly decomposes the sound field using signals obtained from a coincident microphone array. This decomposition both localises the audio objects and enhances them, thereby improving the intelligibility of speech signals. Detailed testing has been carried out in order to quantify the localisation, separation, speech intelligibility improvement and automatic speech recognition performances of the system. Multiple speech sources in both isotropic and directional fluctuating noise under typical application scenarios have been considered.