Robust Continuous Speech Recognition through Combination of Invariant-Feature Based Systems

In the recent years, different types of invariant features have been pro-posed which promise to improve the robustness of speech recognition systems inmismatching training-test conditions with respect to the mean vocal tract lengths.Many state-of-the-art systems make use of system combination. By consideringspeech recognition systems with different front ends, this paper investigates whetherthe system combination of standard-feature and invariant-feature based systemswith ROVER yields improvements in accuracy. Results show that the combina-tion of the considered systems leads to clear accuracy improvements. An erroranalysis also shows that the considered invariant features carry different types ofinformation than the standard ones. The performance achieved with our systemcombination is in the range of what the best systems achieve in literature, althoughour approach does not yet include discriminative training or heteroscedastic featuretransformation.