A brief overview of the main methods of noise estimation

At present, noise estimation has become a key link in speech enhancement technology. In many single-channel speech enhancement algorithms, the real-time noise power spectrum estimation is very important, especially in the case of unknown noise source. The accuracy of the noise estimate directly affects the final effect: if the noise is overestimated, the total self abandonment weak voice will be removed, the speech will be distorted, and if the estimate is too low, there will be more background residual noise.

In recent years, the estimation of noise power spectrum under unsteady environment has been emphasized. The traditional noise estimation method is purely based on the detection of speech activity, which restricts the update of noise in the presence of speech, and the deterioration of reliability performance is obvious when the low input SNR of weak speech signal.
The minimum statistic (minima Statistical,ms) method presented in document "1" is used to estimate the noise by tracking the minimum value of the speech power spectrum within a particular window, multiplied by a coefficient to compensate for the deviation. The variance of the noise estimate obtained by this method is twice times that of the traditional method, and it may occasionally weaken the low-energy Ken, especially when the minimum observation window is very small, and overcomes its limitations at the cost of high computational complexity.

The document "2-3" presents a minimum control recursive averaging (minima controlledrecursive Averaging,mcra) method and an improved MCRA (improved minima controlled Recursive AVERAGING,IMCRA), although they guarantee the accuracy of the noise spectrum estimation, the fixed time window is used to track the minimum value of the smoothed power spectrum of noisy speech, so the estimated noise spectrum has a long delay in the case of noise mutation.

The paper "4" proposes a fast estimation method whose noise spectrum is not dependent on the window length of the same fixed time in the continuous question, but the noise estimation slows down when the voice or noise energy is too high, and when the time is greater than 0. 5 S, it weakens some speech energy.

Note: Shu Wenhao. Research of speech enhancement algorithm based on noise estimation [D]. Shanghai: East China University.

This is a doctoral dissertation, the main methods and principles of noise estimation are described in detail.

Reference documents:

"1" Martin R. Noise Power Spectral Density estimation Based on Optimal smoothing and Minimum statistics[j]. IEEE Trans. On Speech Andaudio processing,2001,9 (5): 504-512.
"2" Cohen I,bergdugo B. Noise estimation by minima controlled Recursive for robust Speech enhancement[j]. IEEE Signal processing letters,2002,9 (1): 12-15.
"3" Cohen I. Noise spectral estimation in adverse environments:improved minima controlled Recursive averaging[j]. IEEE transactions on Speech and Audio processing,2003,11 (5): 466475.
"4" Rangachari S,loizou P C. A noise-estimation algorithm for highly non-stationary Environments[c]//proc. of ICASSP ' 04. Dallas,usa:[s. N ],2004:220-231.