Speaker Diarization

Abstract

We live in the century, where information is strength. Information can be saved in the form of text/multimedia. One such form is Speech audio files. To access those file easily and efficiently, Speaker Diarization is the best option. Speaker Diarization is all about \who spoke when?". This system takes in the input in form of an audio file. The diarization sytem makes the segments of speech file such that each segment is ho- mogeneous. Here, homogeneous segment means that those region of speech contains speech from only one speaker. These segments will be given to the clustering module. This module will merge all the identical segments. To identify the segments from the same speaker, system has to build models for each speaker. These models are constructed using GMMs,using EM algorithm. Feature extraction techniques are ap- plied for extracting speaker specific information. VOice activity detection algorithms are used to differentiate between speech/non-speech reginons and identifying speaker change points. Diarization Systems can be used in Movie analysis, Automatic speech segmentation, Rich transcription, Audio archiving and monitoring, Audio indexing and retrieval. This system doesn't know the number of speakers involved in the sys- tem in advance. We just know is the domain of audio file(which type of recording is this, i.e. telephone conversation, meeting room conversation etc.). The evalua- tion measure used for speaker diarization systems is Diarization Error Rate. This is computed using, Miss, false alarm and confusion. The ideal output expected using diarization system is the speech regions related to speakers, and the speaker lables.