Automatic estimation of position and orientation of an acoustic source by a microphone array network

Abstract

A method which automatically provides the position and orientation of a directional acoustic source in an enclosed environment is proposed. In this method, different combinations of the estimated parameters from the received signals and the microphone positions of each array are used as inputs to the artificial neural network (ANN). The estimated parameters are composed of time delay estimates (TDEs), source position estimates, distance estimates, and energy features. The outputs of the ANN are the source orientation (one out of four possible orientations shifted by 90 degrees and either the best array which is defined as the nearest to the source) or the source position in two dimensional/three dimensional (2D/3D) space. This paper studies the position and orientation estimation performances of the ANN for different input/output combinations (and different numbers of hidden units). The best combination of parameters (TDEs and microphone positions) yields 21.8% reduction in the average position error compared to the following baselines and a correct orientation ratio greater than 99%. Position localization baselines consist of a time delay of arrival based method with an average position error of 34.1 cm and the steered response power with phase transform method with an average position error of 29.8 cm in 3D space.

Distant Speech Recognition Using a Microphone Array Network

“…Using the estimated orientation, the proposed system explic- † Under the conditions in [12], the two-stage ANN and onestage ANN respectively yielded 99.5% and 99.4% for the correct orientation ratio, and 23.2 (20.5) and 25.0 (20.5) cm in 3D (2D) space for the average position error.…”

“…By employing an ANN, it is expected that combining the parameters estimated by the recorded signals of every array of the network will be sufficient to predict the source orientation and give a more accurate source position than that estimated by a single array. The estimation of the position and orientation was performed with a different ANN configuration from that used in [8]. The modifications consist of reducing the two-stage ANN in [8] (by first estimating the orientation and then using this information to select a previously trained ANN to estimate the position) to a one-stage ANN (that can estimate the position and orientation simultaneously), and estimating the orientation of the speaker in eight different orientations equally shifted by 45…”

“…An ANN is used for the position and orientation estimation [8]. ANNs have the ability to learn from and adapt to certain conditions, and can model assumptions about nonlinear/complex physical phenomena, which cannot be solved analytically, responsible for the generation of the given input data of a given process.…”

“…In this work, a robust version of the GCC-PHAT function is used for time delay estimation [11], [12]. The proposed method consists of two basic steps; in the first, a binary mask is used to select only high-energy bands in the cross-power spectrum function in the frequency domain (which corresponds to a more prominent peak in the generalized cross-correlation function in the time domain), and in the second step, only frames that yield physically possible delays are combined to generated a robust crosscorrelation function.…”

“…Spatial information is defined here as the position (location) and orientation (facing angle) of an acoustic source in an enclosed space. The method proposed in [8], which relies on an artificial neural network (ANN), is employed. In this method, a set of input features describing the position and orientation of the source is used as a direct mapping to the true source position and true source orientation.…”