Flexible, robust, and efficient human speech processing versus presentday speech technology

Present-day speech technology systems try to perform equally well or preferably even better than humans under specific conditions. For more complex tasks machines frequently show degraded performance, because their flexibility, robustness and efficiency is lower than that of humans. In order to better understand the system limitations and perhaps further improve system performance, one can try to learn from human behavior and imitate its functionality, without plain duplication. This paper discusses a number of characteristics of human speech processing and compares these with system performance. It is argued that phonetic sciences and speech technology can mutually benefit from each other if they use similar data and similar represen-tations. R. Moore [25] used for this approach the appropriate term Computational Phonetics. 1