Multiple likelihood ratio test-based voice activity detection robust to impact noise in a car environment

Abstract

Since road conditions such as bumps and unpaved roads may cause noise from impacts with a car body, it is hard to accurately detect voice commands for speech-enabled services in a car. To overcome this problem, we propose a voice activity detection (VAD) method that is robust against vehicle body impact noise. The proposed VAD method detects a speech frame on the basis of a multiple likelihood ratio test (LRT) using statistical models of microphone and accelerometer signals, where an accelerometer is used to detect impact noise activity emanating from the vehicle body. The performance of the proposed VAD method is evaluated using a speech corpus recorded in a car moving at a velocity of 30-50 km/h with impact noise ranging from -3 to 1 dB. It is shown from the experiment that the proposed method relatively reduces the average detection error rate by 36.85%, compared to a conventional method using a microphone alone. ©2013 International Information Institute.