Abrupt noise cancellation and speech restoration for speech enhancement

Abstract

In this thesis, we suggest a speech enhancement technique for abrupt noise corrupting speech. Speech quality is improved by removing abrupt noise intervals firstly and then replacing the gaps with appropriate estimates of the prior and posterior speech signal. For improvement of noise detection, we propose the difference residual detective signal which is computed by subtracting the residual signal from modified residual one by using previous frame’s LP coefficients. The threshold detects large pulses in the detective signal. The UVS information of four frames around the detected abrupt noise position is utilized to prevent errors such as plosive speech-to-abrupt noise misclassification. The abrupt noise duration is estimated by the spectral difference between two frames which are 20 msec apart. After removing estimated noise intervals, we applied both one-sided and two-sided linear predictive restoration. To prove the validity of our algorithms, we carry out the detection error tests, the LPC spectral distortion test as the objective test and the PESQ test as the subjective one. Also the recognition test is executed and the results show that the speech quality is fairly well improved.