Applying the Aurora feature extraction schemes to a phoneme based recognition task

The robustness of the ETSI (European Telecommunication Standards Institute) standardized feature extraction schemes is investigated for phoneme based recognition tasks of German speech data. The recognition tasks are an isolated command word recognition and the recognition of connected digits. The motivation of this work is the easy extensibility of a whole word recognition system by allowing also the recognition of phoneme based word HMMs (Hidden Markov Models). The recognition performance has been determined for different numbers of HMM states and different numbers of Gaussians per state. It turns out that fairly high recognition rates can be achieved also for noisy data when applying the second robust ETSI frontend.