Abstract: Today information extraction plays a significant role in management of massive data quantities for different purposes. One of the open challenges in this field is the automatic extraction of information from audio streams. This paper describes a useful metadata extraction system which performs a powerful combination of speech and speaker recognition tasks. The system carries out the speech transcription through a Catalan language recognizer based on Hidden Markov (HMM) tied-state crossword triphones acoustic models, Mel Frequency Cepstral Coding (MFCC) and N-gram language modeling. In addition, a speaker diarization is performed using HMM based segmentation and Perceptual Linear Prediction (PLP) feature extraction. Both speech-to-text transcription and speaker diarization can be utilized as annotation data for multimedia content. In order to make indexing and retrieval more flexible and efficient, the extracted metadata is stored using the MPEG-7 multimedia content description interface. The system has been successfully tested on the recordings of the plenary sessions of the Catalan Parliament.
Index Terms: Metadata extraction, Automatic speech recognition, Speaker diarization, HMM, MPEG-7.