Automatic height estimation from speech in real-world setup

We propose a Gaussian process based regression scheme that provides a direct estimation of the height of unknown speakers and is applicable to real-world autonomous surveillance applications. This scheme relies on utterance-level speech parameterization followed by regression modelling, which estimates the height of the speaker and the uncertainty interval of that estimation. Experiments on the TIMIT database demonstrated that a feature vector composed of the top-50 ranked parameters offers a good trade-off between computational demands and accuracy. The proposed scheme for automatic height estimation was evaluated in the smart-home and public security scenarios offered by the PROMETHEUS database. The averaged relative error of height estimation remained approximately 3%, in both indoor and outdoor conditions, which indicates the good robustness of the proposed scheme.