Method and apparatus for robust noise estimation for speech enhancement in variable noise conditions

摘要

Method for removing embedded acoustic noise and improving speech by identifying and estimating noise in variable noise conditions, the method having the following:the use of a speech/noise classifier togenerate predictive coding coefficient sets that model an incoming signal frame with a higher order LPC and a lower order LPC;the use of the speech/noise classifier to calculate a logarithmic spectral distance between the higher order and lower order LPC resulting from the signal frame, where the logarithmic spectral distance is calculated by two cepstral coefficient sets derived from the higher and lower order LPC coefficient sets;using a speech/noise classifier to compare distance and its short-term motion sequence against a set of thresholds to determine whether the signal frame is speech or noise; updating the thresholds used for the speech/noise classifier based on classification statistics and/or in coordination with other speech activity detection methods;generating multiple linear prediction coding (LPC) coefficient sets as online generated noise models during runtime, each set of LPC coefficients representing corresponding noises, the noise models undercreate and update conditions in which the current signal frame is classified as noise by a probability of the presence of speech and/or the LPC speech/noise classifier;the use of a separate but parallel speech/noise classifier based on assessing the distance between the LPC coefficients of the input signal and the noise models represented by LPC coefficient sets;if the distance assessed is below a threshold, the signal is classified as noise, otherwise the signal is classified as speech;using a noise suppression method by using the probability of the presence of speech to perform noise elimination when ambient noise is stationary;the use of a second noise suppressor having an LPC-based noise/speech classification to enhance noise estimation and attenuation when ambient noise is transient or non-stationary;where the noise estimation by the second noise suppressor takes into account the probability of the presence of speech and adjusts the noise PSD in the frequency range accordingly wherever the conventional noise estimation is insufficient; andusing the newly calculated noise PSD from the advanced noise classification/estimation to produce a refined set of noise suppression enhancements in the frequency range.