Abstract:
Automatic speech recognizers which were once considered as "a dream of mad scientists" have shown considerable success in the last decade. What has made this success possible has been the use of sophisticated mathematical tools along with speech knowledge at various levels. Future success seems to depend on the exhaustive use of the latter. This thesis is an attempt at using prosodic information, which conveys speech knowledge at various levels, in recognition systems. Programs have been developed to extract physical correlates pf prosodic features from the speech signal. Results of analyses with Turkish words and sentences point out some methods to detect linguistic cues from the speech signal. Based on these results, strategies are outlined for a Turkish speech recognizer. Some of these are integrated in an isolated word recognizer and improvements are obtained.
