Results of intership at Lehrstuhl für Multimediakommunikation und Signalverarbeitung (LMS), F.A.U Erlangen
Akshaya Thippur Sridatta
REverberation MOdeling for Speech recognition (REMOS) is an ASR system developed at LMS, which uses constraints from statistical models of various reverberation environments and clean speech training cues to hence solve a 3-D non-linear optimization problem to recognize the novel speech signal provided. This however works in the scenarios of clean speech in reverberant environments with no other interferences. This
model was extended to work in scenarios involving additive background noise/ babble noise by extending the dimensionality of the optimization problem. The main tasks involved modeling the non linear optimization problem and its constraints into piecewise linear components so that they could be solved in real time.