Rasta filtering

RASTA-filtering and Mean Subtraction was originally introduced in connection with the Perceptual Linear Prediction(PLP) type of preprocessing; i.e. bandpass-filtering in the log spectral domain. Slow channel variations should thus in principle be removed. This filtering principle has also been applied to cepstral feature based preprocessing in both the log spectral and the cepstral domains.

A general RASTA filter is defined by

$T(z)\; =\; (\; k\; *\; sum\; (n-(N-1)\; /\; 2)\; *\; z^\{-n\})\; /\; (1-\; ho/x)\; ,!$

where the numerator is a regression filter of odd order N and the denominator is a leaky integrator. The pole controls the lower cut-off frequency and is normally placed in the neighbourhood of 0.9.

A simple variant of RASTA-filtering is mean subtraction which corresponds to a moving average filter. Filtering is normally performed in the cepstral domain (CMS). The mean corresponds to the long term cepstrum and is normally computed on the speech part utterance by utterance. A silence/speech-detector is thus necessary.