A database for microphone array experimentation

Figure 3: Example of phase correlation between two microphones. The peak of this function indicates the inter-channel delay. index associated with peak value of f(t). This delay estimator is computationally convenient and more robust to noise and reverberation than other approaches based on cross-correlation or adaptive ltering. In ideal conditions, the output of Equation (5) is a delta function centered on the correct delay. In real applications with a wide band signal, e.g., a speech signal, the outcome is not a perfect delta function. Rather it resembles a correlation function of a random process. The time index associated with the maximum value of the output of Equation (5) provides an estimation of the delay. The system can produce wrong answers when two or more peaks of similar amplitude are present, i.e., in highly reverber-ant conditions. The resolution in delay estimation is limited in discrete systems by the sampling frequency. In order to increase the accuracy, oversampling can be applied in the neighborhood of the peak, to achieve sub-sample precision. Fig. 3 demonstrates an example of the result of a cross-power spectrum time delay estimator. Once the relative delays associated with all considered microphone pairs are known, the source position (x s ; y s) is estimated as the point that would produce the most similar delay values to the observed ones. This optimization is performed by a downhill sim-plex algorithm 6] applied to minimize the Euclidean distance between M observed delays ^ i and the corresponding M theoretical delays i : An analysis of the impulse responses associated with all the microphones, given an acoustic source emitting at a speciic position, has shown that constructive interference phenomena occur in the presence of signiicant reverberation. In some cases, the direct wavefront happens to be weaker than a coincidence of reeections, inducing a wrong estimation of the arrival direction and leading to an incorrect result. Selecting only microphone pairs that show the highest peaks of phase correlation generally alleviates this problem. Location results obtained with this strategy show comparable performance (mean posi-Reverb. Time Average Error 10 mic pairs 4 mic pairs 0.1sec 38.4 cm 29.8 cm 0.6sec 51.3 cm 32.1 cm 1.7sec 105.0 cm 46.4 cm Table 1: Average location error using either all 10 pairs or 4 pairs of microphones. Three reverberation time conditions are considered. tion error of about 0.3 m) at reverberation times of 0.1 s and 0.6 s. …