Robust Speech Detection for Noisy Environments

Abstract

This paper presents a robust voice activity detector (VAD) based on hidden Markov models (HMM) to improve speech recognition systems in stationary and non-stationary noise environments: inside motor vehicles (like cars or planes) or inside buildings close to high traffic places (like in a control tower for air traffic control (ATC)). In these environments, there is a high stationary noise level caused by vehicle motors and additionally, there could be people speaking at certain distance from the main speaker producing non-stationary noise. The VAD presented in this paper is characterized by a new front-end and a noise level adaptation process that increases significantly the VAD robustness for different signal to noise ratios (SNRs). The feature vector used by the VAD includes the most relevant Mel Frequency Cepstral Coefficients (MFCC), normalized log energy and delta log energy. The proposed VAD has been evaluated and compared to other well-known VADs using three databases containing different noise conditions: speech in clean environments (SNRs mayor que 20 dB), speech recorded in stationary noise environments (inside or close to motor vehicles), and finally, speech in non stationary environments (including noise from bars, television and far-field speakers). In the three cases, the detection error obtained with the proposed VAD is the lowest for all SNRs compared to Acero¿s VAD (reference of this work) and other well-known VADs like AMR, AURORA or G729 annex b.