In-vehicle speaker recognition using independent vector analysis

As part of human-centered driver assist framework for holistic multimodal sensing, we present an evaluation of independent vector analysis for speaker recognition task inside an automotive vehicle. Independent component analysis-based blind source separation algorithms have attracted attentions in recent years in the application of speech separation and enhancement. Compared to the traditional beamforming technique, the blind source separation method may typically require less number of microphones and perform better under reverberant environment. We recorded two speakers in the driver and front-passenger seats talking simultaneously inside a car and used independent vector analysis to separate the two speech signals. In the speaker recognition task, we show that by training the model with the speech signals from the IVA process, our system is able to achieve 95 % accuracy from a 1-second speech segment.