Speech Enhancement in Car Noise Environment

Abstract

Minimum Mean Square Error estimator in log spectral (LogMMSE) was evaluated in 19 noisy scenarios within 5 categories of noise (vehicle, street and babble) to examine the noise suppression ability of this algorithm compared to Minimum Mean Square Error Short Time Spectral Amplitude (MMSE STSA) with or without Speech presence Uncertainty detection (MMSE w/SPU) [1]. The aim of this evaluation is to evaluate the performance of LogMMSE algorithm operating in specific noise condition. Noise contaminated sentences, ranging from 0dB to 20dB SNR, were imported to MATLAB to simulate the noise suppression ability of algorithm in different vehicle noise environments. Result of the enhanced speech was scored by subjective listening test; the objective PESQ (ITU-T P.862) and spectrogram were used to support the performance of algorithms. Emphasis was placed on the result which can be used to interpret the suitable algorithm might be used in car noise environment. The results form this study suggested that LogMMSE could be suitable in car noise environment.