Room localization for distant speech recognition
Juan A. Morales-Cordovilla, Hannes Pessentheiner, Martín Hagmüller, Gernot Kubin · 2014
The problem of room localization is to determine where, in a multi-room environment, a person is producing a speech ut-terance. In our work, we are exploiting the information gained from a network of microphones installed all over a house, where the lack of calibration of the microphone energies creates an ad-ditional challenge. This paper compares room localizers based on different features (such as energy and cross-correlation be-tween microphones) and classifiers (such as neural networks and discriminative analysis). In order to evaluate the differ-ent room localizers in terms of word accuracy this paper also presents a complete distant speech recognition system which tries to take advantage of synergy between the different compo-nents without using any oracle information. Finally, the system is analyzed in terms of computational and time resources. Index Terms: Distant speech recognition; microphone net-work; VAD; room localization; machine learning classification; enhancement; reverberant and noisy environment. 1.