An effective missing feature compensation method for speech recognition at noisy environment

Abstract

It is a challenge task for maintaining high correct word accuracy rate (WAR) for state-of-art automatic speech recognition (ASR) systems when the SNR goes very low. To deal with such situation, the missing feature technology (MFT) has shown as one of the mainstream algorithms. In principle, conventional MFT firstly separate the unreliable spectral bins from the reliable ones. Then the unreliable bins are reconstructed by missing feature algorithm [7]. When SNR goes low, the performance of the conventional MFT for ASR system is limited since both the reliable and unreliable spectral bins will be corrupted by the noise components. In this paper, a novel missing feature compensation method was developed by considering compensating both unreliable and reliable spectral bins. With the assumption of GMM distribution of the clean speech spectral vector, a dual MFT (DMFT) algorithm is developed, where the reliable spectral bins corrupted by noise have been compensated by removing the noise components. Several experiments have been carried out to evaluate the performance of the proposed DMFT algorithm by using AURORA2 database. From the results, it is clear to see that the proposed DMFT algorithm improves the WAR under all types of noises at different SNR levels compared with the traditional MFT algorithm.