In most speech enhancement algorithms, it's  assumed that an estimate of the noise spectrum is avaible, knowing that this estimate is critical in for the perfomance of the speech-enhancement algorithms such as:

  • Wiener Filter in the wiener algorithms (Lim and Oppenheim, 1978).

and so on.

As the noise estimation has a great impact in speech-enhancement process, in one hand if the noise estimate is too low, annoying residual noise will be audible in the other hand if the noise estimate is too high, signal will be distorted resulting to intelligibility loss.

The easiest and simplest way of estimating noise spectrum is during silent or pause using voice activity-detection algorithm (VAD) unfortunately this way of doing has a major backwards, it assumes that the noise is stationnary which is not the case in real situation where noise noise spectrum change continuously over time therefore other approches should be used so that noise estimation is updated over time regard less to signal presence.

Many noise-estimation algorithms have been proposed for speech enhancement applications where the most known are those from:

  • Martin in 2001.
  • Cohen in 2002.
  • Lin and Al in 2003.
  • Rangachari and Al 2004.
  • Rangachari and Philipos C. Loizou in 2005

Unfortunaltely all of those algorithms won't be discussed here, in the other side we will présent our approch witch estimate the noise standard déviation with unknown distribution of the signals occurences also known as Algorithms and applications for Etimating the standard Deviation of additive white Gaussian Noise when Observations are not signal-Free presented on september 2007 by Dominique Pastor and Asmaa Amehraye. A robust estimation of noise standard deviation with unknown distribution of signals occurrences


jvcounterThis Week105
jvcounterLast Week929
jvcounterThis Month1932
jvcounterLast Month2274

There are 10 visitors online.
10 Guests