Reverberation Mitigation of Speech

Speech data corrupted by reverberation, especially in combination with noise, has a drastically negative effect on the performance of speaker identification algorithms. Current approaches to the removal of reverb and noise rely on specific knowledge of room characteristics causing the reverb.The team of In-Depth Engineering and the University of Maryland propose to apply an innovative, generalized, adaptive algorithm for reverberation removal that is based on a model of human sound analysis that occurs in the primary auditory cortex. The Phase I feasibility study showed that the model can be used to provide a robust technique for removing both reverb and noise from speech data. The team proposes to complete the experimentation and analysis initiated in the Phase I effort to optimize the application of the cortical model, making use of real-world reverberated and noisy data from the Air Force, and testing with a speaker identification algorithm (SID).The use of the cortical features to directly generate iVectors to feed the SID will also be explored, and a real-time processing version of the cortical algorithms will be developed.Speech data corrupted by reverberation, especially in combination with noise, has a drastically negative effect on the performance of speaker identification algorithms. Current approaches to the removal of reverb and noise rely on specific knowledge of room characteristics causing the reverb.The team of In-Depth Engineering and the University of Maryland propose to apply an innovative, generalized, adaptive algorithm for reverberation removal that is based on a model of human sound analysis that occurs in the primary auditory cortex. The Phase I feasibility study showed that the model can be used to provide a robust technique for removing both reverb and noise from speech data. The team proposes to complete the experimentation and analysis initiated in the Phase I effort to optimize the application of the cortical model, making use of real-world reverberated and noisy data from the Air Force, and testing with a speaker identification algorithm (SID).The use of the cortical features to directly generate iVectors to feed the SID will also be explored, and a real-time processing version of the cortical algorithms will be developed.