METHOD AND APPARATUS USING HARMONIC MODEL BASED FRONT END FOR SPEECH RECOGNITION

摘要

PURPOSE: A method and apparatus using a harmonic model based front end for speech recognition are provided to decompose a speech signal into a harmonic component and a random component and sum up the harmonic component and the random component to form a noise-reduced feature vector, to thereby reduce a noise in the speech signal. CONSTITUTION: A part of a noisy speech signal is decomposed into a harmonic component and a random component(304). A scaling parameter for at least the harmonic component is determined(306). The harmonic component is multiplied by the scaling parameter for the harmonic component to form a scaled harmonic component. The random component is multiplied by a scaling parameter for the random component to form a scaled random component. The scaled harmonic component and the scaled random component are summed up to form a noise-reduced value(310).