ANOMALY DETECTION IN ACOUSTIC WITH GAUSSIAN MIXTURE MODEL

• Main Author: PUTRI (NIM: 10114006), ZULFARIDA
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/31917
• Institution: Institut Teknologi Bandung

Speech processing is one of sub field in signal processing that enables computer to receive an input of spoken speech. In speech processing, it is important to differ-entiate which part of audio signal that contains speech. Therefore, speech activity detection is used as pre-processing step in speech processing. However, detecting which part of the signal that contains speech is hard problem due to increasing noise in the background can decrease the effectiveness of speech detecting system. This non speech voice is called anomaly. The problem of detecting anomaly in acoustic can be seen as statistical problem. In this essay will be discused the steps to estimate the distribution of normal speech. Normal speech distribution will be modeled by Gaussian Mixture Model because this distribution can model various distribution that has several modes and it is often used in model of human biometric system. The parameter of Gaussian Mixture Model will be estimated with Expectation-Maximization algorithm. In modelling the normal speech distribution, the data that will be used is two features of syllable rate. After estimating the Gaussian Mixture Model, a threshold will be determined for the speech to be categorized as speech or anomaly. To determine this threshold speech and anomaly data will be used as the test data.