|Title||Automatic Speech Recognition and Transcription for a topic-related Segmentation Tool in Audio Streams|
|Tutor||Dipl.-Ing. Florian Kaiser|
|Professor||Dr.-Ing. Thomas Sikora|
|Abstract||In this master-thesis the interest lies in extracting the structure of spoken audio streams with means of speech transcription. The main focus is put on News data that can be structured according to diverse types of information; speech, music, particular speakers, advertisement or other topics. |
This thesis uses the Term Frequency-Inverse Document Frequency (TF-IDF) method in text index to develop a topic detection system based on the news from ABC World News. 15 vocabularies are designed for the topic detection system, aiming at detecting the topics of the news reports in text format, with a successful recognition rate up to 80%. Performance of the topic detection system is evaluated with common tools such as the truth table, confusion matrix and single topic precision, recall and F-measure values.
|Key words||speaker diarization (SD), speaker clustering (SC), automatic speech recognition (ASR), topic detection and tracking (TDT), term frequency-inverse document frequency (TF-IDF)|