Robust continuous speech recognition through combination of invariant-feature based systems

Authors: Florian Müller, Alfred Mertins

Abstract:

In the recent years, different types of invariant features have been proposed which promise to improve the robustness of speech recognition systems in mismatching training-test conditions with respect to the mean vocal tract lengths. Many state-of-the-art systems make use of system combination. By considering speech recognition systems with different front ends, this paper investigates whether the system combination of standard-feature and invariant-feature based systems with ROVER yields improvements in accuracy. Results show that the combination of the considered systems leads to clear accuracy improvements. An error analysis also shows that the considered invariant features carry different types of information than the standard ones. The performance achieved with our system combination is in the range of what the best systems achieve in literature, although our approach does not yet include discriminative training or heteroscedastic feature transformation.