13 Voice Activity Detection for In-Car Communication Systems

Abstract

Occupants in a driving car are often confronted with severe noise levels that impede dialogue among them. Particularly passengers seated at different rows must raise their voices inconveniently and shout against the noise to bridge the gap. In this situation, in-car communication (ICC) systems may assist: these systems pick up speech using microphones close to different speakers and - with only little delay - reinforce the signal via loudspeakers at the listeners’ positions. Signal processing algorithms in these systems aim at enhancing the audio reception of the listeners by amplifying speech and reducing the impact of background noise. Many of these algorithms require accurate and robust estimates of speech and background characteristics to ensure that only desired speech is reproduced whereas noise and other undesired components are kept unamplified. For this purpose, voice activity detection (VAD) is a pivotal factor: noise suppression just as many other algorithms rely on the detection of speech components that must be captured and preserved. In this chapter, the concept of VAD is discussed subject to particularly challenging demands on signal processing imposed by low-latency requirements that are typical for an ICC system but that can be also found in other applications such as hearing aids. Features for VAD are derived that can cope with a low spectral resolution resulting from very short time frames that are chosen in favor of low latency. The subsequent evaluation further explicitly addresses the delay introduced by VAD and its impact on different algorithms within an ICC system.