Coordinated training of noise removing networks

When a recognition system designed with a noise-free speech signal is deployed in the real world, it experiences severe degradation due to environmental acoustic ambient noise. An enhancement technique based on a noise removing network (NRN) which has recurrent connections is proposed. This NRN is trained in coordination with a time delay neural network (TDNN) which has been trained for noise-free speech recognition tasks. The proposed enhancement technique has favorable performance for white/colored Gaussian noisy speech when compared with the NRN without coordinated training and the classical linear Wiener filtering methods.<>