Real-time Speaker Diarization on TI OMAP3530

In this paper a real-time implementation of a Speaker Diarization algorithm on TI OMAP3530 is discussed. The addressed algorithm is made of two main stages: the acoustic feature extraction and the speech activity classifier. Both of them have been implemented and run on the ARM Cortex-A8 processor of the OMAP3530 chip. Classifier training has been performed offline on a PC by using a suitable speech database (i.e. the AMI Corpus). The acoustic feature extraction process represents the bottleneck in terms of computational burden, that is why the scalability issue has been faced and code optimization (using ARM NEON architecture) adopted to make the algorithm working in real-time. Experimental results prove the effectiveness of the approach.