Output details

Multichannel Online Blind Speech Dereverberation with Marginalization of Static Observation Parameters in a Rao-Blackwellized Particle Filter

This invited paper presents several novel multi-channel online blind dereverberation algorithms. These sequential algorithms facilitate implementation of real-time applications, and are important for human-machine-interaction. The algorithms utilise flexible speech models and provide a major advance on earlier Bayesian batch-mode techniques (BMT). The results show a speech-to-reverberation ratio improvement of 12dB@T60=0.45seconds for a single sensor, with a further 4dB improvement for 10 sensors, far outperforming BMT. The techniques formed the basis for joint enhancement/localisation techniques in EPSRC/DSTL EP/H012699/1 (£117k) and DSTL/UDRC EP/K014277/1 (£4.3M). The results are an extension of work from EP/D051207/1 and two IEEE Conferences (DoI:10.1109/ISCAS.2008.4542145, DoI:10.1109/ICASSP.2008.4518680).