Recognition of Regional Variants of German using Prosodic Features

Diplomarbeit (Word 97-Datei, gezippt 3.45 MB), Mai 2001.

This thesis attempts to recognize national variants of German in Austria and Germany using only prosodic features. 
An introduction to speech processing with special consideration of speech recognition and language identification is given. Fundamental frequency (F0) and Intensity of the speech signal are analyzed. Parameterization of F0 (Fujisaki, Intofit) is carried out and statistical features (Standard deviation, skewness, kurtosis, percentiles) are calculated from signals such as derivative and correlation of F0 and Intensity.
The features are evaluated using the t-test and a simple classification algorithm using combinations f up to three features. Combinations with Fujisaki parameters yield the best results with recognition rates of 72%.
The small size of the data-corpus is a drawback of the study.