Optimizing DNN-HMM-based ASR system using speech enhancement techniques

In this project, the final purpose is to utilize DNN-HMM-based automatic speech recognition (ASR) system to record the conversations and provide transcriptions for customers. Although current ASR algorithms have achieved the satisfactory performance in a quiet environment, their performances degrade significantly in real environments due to the background noise, interfering speakers and reverberation. This problem is the so-called cocktail party problem. Thus, the purpose of this project is to apply speech enhancement techniques to optimize the DNN-HMM-based ASR system so that acquire the lower word error rate (WER) in complex noisy environment.