A robust feature extraction for automatic speech recognition in noisy environments

This paper presents a method for extraction of speech robust features when the external noise is additive and has white noise characteristics. The process consists of a short time power normalisation which goal is to preserve as much as possible, the speech features against noise. The proposed normalisation will be optimal if the corrupted process has, as the noise process white noise characteristics. With optimal normalisation we can mean that the corrupting noise does not change at all the means of the observed vectors of the corrupted process. As most of the speech energy is contained in a relatively small frequency band being most of the band composed by noise or noise-like power, this normalisation process can still capture most of the noise distortions. For signal to noise ratio greater than 5 dB the results show that for stationary white noise, the normalisation process where the noise characteristics are ignored at the test phase, outperforms the conventional Markov models composition where the noise is known. If the noise is known, a reasonable approximation of the inverted system can be easily obtained performing noise compensation still increasing the recogniser performance.