The Low SNR Speech Enhancement Method Based on Multi-Taper Spectrum Estimation and Spectral Subtraction of Geometric

The traditional spectral subtraction methods tend to produce "music noise" under low signal-to-noise ratio (SNR) due to the independence assumption of speech and noise. The intelligibility of speech is reduced greatly. In this paper, we proposed a speech enhancement method via integrating multi-taper spectrum (MTM) estimation and spectral subtraction of geometric (GA). Our method employs the MTM to estimate the power spectrum of the noisy speech, and the improved minima controlled recursive average method to track the estimated noise spectrum in real time. Furthermore, the GA is used to calculate the gain function and speech signal is recovered accurately. Our method can reduce the spectral distance between the noisy speech and the clean speech under the low signal-to-noise ratio prominently, and restrain the background noise of the enhanced speech effectively. Experiments on the IEEE dataset, PESQ and LSD as the evaluation metrics, show that our method improves the speech intelligibility significantly.