Speech enhancement method using self-adaptive time-shift and threshold discrete cosine transform

In view of the limitation of the existing speech enhancement methods under a low SNR, this paper proposes a speech enhancement method using self-adaptive time-shift and threshold discrete cosine transform. First, with the improved soft-threshold function to deal with the discrete cosine transform coefficients, we can not only eliminate the noise of noise-dominant frames, but also eliminate the noise of signal-dominant frames; the threshold is also selected self-adaptively based on the SNR, which can largely retain the original characteristics of the speech.Secondly, the shift of the analysis window is self-adapted according to the pitch period, reducing the white noise of the fixed window-shift.And a weighted autocorrelation function is introduced for pitch detection combined by the short-time autocorrelation function and the short-time average magnitude separation function, improving the precision of pitch detection and robustness to noise. Theoretical analysis and simulation results show that the output SNR of this method has increased greatly and the robustness to noise is better when the input SNR is as low as -5dB, compared with the empirical mode decomposition algorithm and the subspace algorithm.