Assessment of clinical voice parameters and parameter reduction using supervised learning approaches

Over the years, new measurement techniques and procedures for data collection have been introduced in voice research. Different signals can be measured and on basis of these signals a large number of parameters can be calculated describing different aspects of voice production and quality.
However, since parameters were not introduced in an organized manner, but are rather "grown naturally", analysis on their interdependencies and possible vulnerabilities towards infuencing factors has been neglected. If parameters are mathematically dependent they cover the same information. Therefore only one of a group of redundant parameters is needed for data interpretation. Further, if parameters are vulnerable towards certain infuencing factors they may not be applicable in some research settings. Results may be distorted by these infuencing factors. Also the large number of parameters in use hinders study comparability. Especially in clinical settings a small number of robust parameters is desired, since in these settings different influencing factors that may affect parameters can usually not be avoided. Therefore in this thesis a large scaled investigation of parameters describing voice characteristics and their interrelations is conducted.