Cascaded speech enhancement system using deep learning method

Abstract

Here, a two-stage cascaded noise minimization from noisy speech is proposed for noise cancellation from highly corrupted speech signals. In the first stage, corrupted speech is passed through speech enhancement system based on wavelet domain adaptive filter using least mean square algorithm (WDAF-LMS) and performance is evaluated for noisy signal corrupted by babble noise, car noise and machine gun noises. Then this output is given to second stage for further improvement. This is fully connected deep neural network using stochastic gradient descent with momentum optimizer (FCDNN-SGDM) used to improve the quality of speech signal. The system is tested for highly corrupted noisy speech signals where noise signal power level is equal to or more than clean signal power. Input signal-to-noise ratio (SNR) level is taken as 0 dB and -5 to -13 dB. The proposed system improved the quality and intelligibility of speech at all SNR levels for all three noises.