Name spotting over low signal-to-noise ratio (SNR) using Blind Source Separation and Connectionist Temporal Classification

摘要

Speech recognition in a signal with low SNR is a very challenging task. When the distance between the mic and the source is large, the mic records a mixture of Speech and Noise. This paper presents a Speech recognition system which performs Blind Source Separation using Degenerate Unmixing Estimation Technique to separate speech from noise. This system uses a Deep Recurrent Neural Network based method in order to achieve robust speech recognition. An experiment comparing the efficiency of the aforementioned system with an already established Speech Recognition system is presented at the end.