COVERT COLLECTION AND AUTOMATED ANALYSIS OF VIBROACOUSTIC INTELLIGENCE FROM DRONE MOUNTED LASER DOPPLER VIBROMETERS

The synthesis of Laser Doppler Vibrometers (LDVs) with autonomous or remotely piloted vehicles such as drones has the potential to enable highly sensitive, non-invasive and discrete vibroacoustic intelligence gathering processes in hostile environments without risk to human life. This work builds upon a previously developed vibroacoustic noise reduction and speaker diarisation system by exploring the effect of feature extraction parameters on diarisation performance. By tuning the Mel Frequency Cepstral Coefficients (MFCC) and x-vector windowing parameters - how many samples are used to produce a single feature vector - the optimal combination was determined to be 0.305 and 0.5 seconds, respectively, resulting in an error of approximately 5%. This work also presents a live or'online' vibroacoustic intelligence processing and analysis system by utilising an open-set clustering algorithm - Real-Time Exponential Filter Clustering (RTEFC). Similarly, the effect of the similarity threshold D and the exponential filter parameter α on diarisation performance was explored. The most effective combination was 0.96 and 0.75, respectively, resulting in an error of approximately 10%. Furthermore, a live transcription stage has also been included using the Microsoft Azure Speech-to-Text API, automating another important intelligence analysis process.