A Robust Speaker Recognition Algorithm Using the Wavelet Transform

摘要

PURPOSE: A system for identifying a speaker strong to an external noise is provided to use a wavelet transform to separate original signals into four subbands, and to construct independent codebooks for three frequency bands having excellent capacities to finally have one decision-making value, so as to prevent a noise of a subband from influencing other subbands. CONSTITUTION: A voice detector detects a voice start point and a voice end point. A voice analyzer analyzes voices of each word, and finally finds a linear prediction coefficient and a mel-frequency ceptrum coefficient. If an algorithm is a vector quantization algorithm, a trainer makes codebooks representing each voice by using a K-means clustering algorithm for specific vectors obtained from the voice analyzer. A recognizer compares inputted speaker data with the codebooks to select a codebook having the nearest vector space distance, and decides a speaker corresponding to the codebook as recognition.