A more effective speech enhancement algorithm under non-stationary noise environment

Aiming at non-stationary noise environment and low SNR (signal to noise ratio), we propose a more effective speech enhancement algorithm. The non-stationary noise estimation is based on noisy speech signal properties from low-frequency regions and high-frequency regions. Section 1 of the full paper explains what we believe to be a more effective algorithm; its core consists of: (1) we construct a time-varying weighting value for each frame of windowed speech signals through non-stationary noise estimation to realize the real time noise estimation; (2) utilizing the human auditory masking properties, we use Eq.(12) to calculate the masking threshold values of each frame of speech signals in its different BARK regions; (3) we then use the calculated masking threshold values to self-adaptively adjust the speech enhancement coefficients calculated by using Eq.(4). Section 2 simulates our speech enhancement algorithm to verify its effectiveness; the simulation results, given in Figs.2 and 3 and Table 1, show preliminarily that our algorithm is more effective for reducing background noise, improving SNR and decreasing speech distortion.