Research Of Transient Noise Suppression Algorithm In Speech

Typical transient noise is an abrupt or impulsive sound followed by a short time decaying oscillations. The existence of transient noise can seriously degrade the speech quality. Due to the widely spread across the frequency domain and completely aliasing with speech signal in the time and frequency domains, traditional speech enhancement algorithm can’t effectively suppress the transient noise, and may cause distortion of the speech signals. In this paper, we present two novel algorithms for transient noise suppression:(1) We propose a transient noise suppression algorithms based on the modified minima controlled recursive averaging(MCRA) and optimally-modified log-spectral amplitude (OM-LSA) estimator. The transient noise is regard as a signal, and the power spectrum distribution(PSD) of slowly varying speech and background noise is estimated by the modified MCRA. Then, we estimate the PSD of transient noise by enhancing transients and attennuating non-transient components using OM-LSA. Finally, the transient and background noise are suppressed by OM-LSA simultaneously.(2) Based on a detection-restoration scheme, we propose an algorithms based on Gammatone frequency cepstral coefficients(GFCC) and speech signal reconstruction. In the stage of transient noise detection, using the human ear auditory perception effect, GFCCs are extracted frame-by-frame. Then the GFCC distinction of transients and speech is used to detect the existence of transient noise. In the stage of transient noise suppression, based on the short-time stationary and correlation of speech, a bidirectional linear prediction algorithm is employed for noisy speech signal construction.The performances of proposed algorithms are evaluated, including the objective test and subjective test of the speech quality. Experimental results show that, compared with two traditional transient noise suppression algorithms, the two proposed algorithms achieve better results of transient noise suppression, and effectively improve both the time domain and frequency domain distortion of the noisy speech. In addition, the subjective auditory feels are also improved.