Comparative Study of Deep Learning Techniques Used for Speech Enhancement

Speech degradation is caused by various reasons such as environmental disturbances captured during the speech signal recording, like reverberation, babble, etc. The objective of speech enhancement algorithms is to reduce noise while improving the overall quality and intelligibility of degraded speech signals and reducing auditory fatigue. Speech enhancement finds its application in various areas that require the need for noise-free clean speech signals, such as, speaker recognition, speech recognition, hearing aids, and almost every communication medium that is based on voice. This paper aims to study the temporal and spectral features of the speech signal which were exploited to create Neural Network models such as CNN, RNN, and LSTM models to perform the task of speech enhancement. Finally, an end-to-end deep learning model, WaveNet+LSTM, which takes raw audio data or spectral magnitude using STFT asinput and maps it to the targeted clean speech and learns by minimizing an energy-conserving loss function was created and analyzed. This model was compared to the other neural network models and the results analyzed. Speaker-specific training of the WaveNet+LSTM modelwas done to compare its performance when trained using frequency domain and time domain features.