Apparatus and Method For Feature Compensation Using Weighted Auto-Regressive Moving Average Filter and Global Cepstral Mean and Variance Normalization

摘要

PURPOSE: A feature compensating device using a weighted auto-regressive moving average filter and a global cepstral mean and variance normalization and a method thereof are provided to use global mean and variance from cepstrum of all data, thereby preventing performance lowering of a voice recognizing system at a remote place. CONSTITUTION: An MFCC(Mel-Frequency Cepstral Coefficients) feature extracting unit(100) extracts a training voice cepstrum and a recognition voice cepstrum from a voice signal of each frame. A cepstrum mean and variance normalizing unit(110) normalizes the training voice cepstrum and the recognition voice cepstrum. A weighted auto regressive moving average filter(130) performs weighted auto-regressive moving average filtering on normalized cepstrum time-series. A voice recognizing unit(160) selects a sentence to maximize likelihood of an HMM(Hidden Markov Model) sound model of a sound model training unit about the recognition voice cepstrum by Viterbi decoding.