D. Van Compernolle, T. Claes, J. Smolders, F. Henderieckx, E. Kesters

The performance of speech recognition systems degrades in unexpected communication environments. Special techniques to increase the robustness of the speech parameters are necessary in noisy environments. Past research has already shown that (Jah-)RASTA improves the performance of isolated word recognition (IWR) systems in the presence of convolutional and additive noise. Filtering the logarithmic spectrum, RASTA suppresses the spectral components that change more slowly or quickly than the typical range of change of speech.

An important disadvantage of RASTA is that it increases the dependence of the data on its previous context. This can degrade the performance of context-independent subword-unit recognizers. But also in the case of continuous word recognition (CWR), words can have an important influence on the words that follow. This effect was an important topic for research. Although this context-influence is real, experiments in car-environments have shown that the gain in performance by using RASTA is still larger than the decrease caused by the context-influence in the case of word models. This influence seems to be more important for phoneme-based recognizers, which use context-independent phonemes.