The aim of the thesis is to compare different techniques for information retrieval in large collections of audio material. More specifically, in the work we compare the use of query methods based on keyword search and methods based on vector word insertions. For the purposes of evaluating different methods of information retrieval, we have developed a search system for audio recordings. Today, high-quality speech recognition systems are also available for the Slovenian language. These enable machine transcription of audio recordings, and the latter provides the possibility of precise searching, as the transcripts are equipped with time stamps of the words that appear in the transcript. In the thesis, we have chosen the RTV radio broadcast database as the domain. The developed query application serves as a demonstration system for searching in large collections of audio recordings. Our work has shown that methods based on word embeddings achieve higher accuracy in content retrieval, as they capture the meaning of words more effectively within their local context. Nevertheless, keyword-based search methods remain faster and computationally less demanding. Combining embedding-based methods with keyword search allows for optimal accuracy in retrieval tasks.
|
