Automatic lecture Transcription in the Greek language

Summary

This thesis focuses on Automatic Speech Recognition (ASR) models applied to Greek-language AI lectures that incorporate technical terminology primarily derived from English. Initially, existing ASR models were evaluated based on their transcription accuracy and error rates. Subsequently, additional data were collected, an existing model was retrained, and new audio samples were analyzed. A key motivation behind this work is to support individuals with hearing impairments, enabling them to attend and comprehend AI lectures through accurate transcriptions. Within this context, the thesis also explores the feasibility of integrating lip-reading capabilities with transcription systems. Given the cognitive challenge of simultaneously reading subtitles and interpreting a speaker's lip movements, it was particularly interesting to examine whether such a combination is practical and how the two modes can complement each other. Data for this study were gathered through custom-designed surveys and targeted field research. The ultimate goal is to develop or enhance an ASR model tailored to the needs of hearing-impaired individuals, improving their ability to follow academic lectures effectively.