Evaluation of root-normalised front-end (RN LFCC) for speech recognition in wireless GSM network environments

With the emergence of mobile telephony, the conditions of use of speech recognisers have been completely reconsidered. Speech can be input from a wide range of environments, having to face several degradation causes principally induced by noise and/or channel distortions. The root-normalised front-end is evaluated on several databases. Special attention is given to the GSM environment where noise, channel and coder effects are analysed. Experimental results indicate that this analysis model, when compared to the more conventional mel-scale cepstral representation (MFCC), is proved to be more robust to several types of perturbations. This advantage is maintained when using explicit channel and noise compensation schemes such as NSS (non-linear spectral subtraction).