In MARVEL, FORTH implements GPURegex, which is the component that performs high-performance processing of captioning data exported from audio collected by the MARVEL pilots. Specifically, after the anonymization procedure, the raw audio data is processed by the MARVEL component, namely Automated Audio Captioning (AAC) technology, owned by Tampere University (TAU). AAC produces the captions as they occur from the audio files. The exported captioning data are then forwarded as input to the GPURegex component. Besides captioning data, GPURegex receives as input the keywords that will be searched in order to indicate an event. Example keywords “siren” and “injury” could point to a post-accident situation, while “cheer” and “applause” could imply a celebration.

GPURegex reports events by enabling fast pattern matching against those captioning data based on the characteristics and capabilities of modern commodity hardware accelerators (e.g., high-end GPUs). GPURegex can support processing on top of different types of hardware devices (i.e., multicore CPUs and integrated GPUs) since it is implemented using the OpenCL programming framework[1] that enables uniform execution across heterogeneous devices. Utilizing hardware accelerators for pattern matching can significantly speed up the process of analyzing and processing large amounts of data. GPUs are designed to handle massive amounts of data computations in parallel, making them ideal for tasks that involve high-performance computing and large data sets. A GPU parallelizes the processing of the data, which means that multiple computations can be performed simultaneously. This can significantly speed up the data analysis and execution time required, enabling more efficient pattern matching.