INFERENCE OF ACOUSTIC SOURCE DIRECTIVITY USING ENVIRONMENT AWARENESS A. Brutti, M. Omologo and P. Svaizer Fondazione Bruno Kessler CIT-irst via Sommarive 18, 38123 Trento, Italy phone: + 39 0461314529, email: brutti@fbk.eu ABSTRACT Acoustic maps derived from the Generalized Cross-Correlation Phase Transform (GCC-PHAT) computed on the signals acquired by a set of distributed microphones can be effectively used for the localization of active acoustic sources. When the microphone pairs surround a given area with a good angular coverage, directional characteristics of the sources can also be inferred, based on the relative amplitudes of the GCC-PHAT peaks and the geometry of propagation in the given environment. This paper presents a novel method for estimating the radiation pattern of an acoustic source which combines GCC-PHAT observations with accurate descrip- tors of the environment characteristics, i.e. reverberation time. Ex- periments on simulated data show that the source emission pattern can be estimated in an effective way under noisy and reverberant conditions. 1. INTRODUCTION Acoustic scene analysis for audio processing takes advantage of an accurate characterization of the sound sources, for instance in terms of spatial positioning and radiation properties. As far as the latter is concerned, in general acoustic sources present emission patterns which are far from being omnidirectional in particular at higher fre- quencies [10]. As an example, Figure 1 sketches the typical radia- tion pattern of humans at two frequencies [5, 7]. 200Hz 1500Hz Figure 1: Rough shape of the head radiation pattern at two frequen- cies. To this regard, the SCENIC project 1 focuses on techniques for achieving acoustic awareness of an environment and on methods for employing the acquired awareness in source characterization and enhancement of the recorded signals. The specific radiation pattern, and consequently the orientation of the source, are cru- cial features in many speech related algorithms which rely on the assumption that direct wave-fronts prevail in the multipath prop- agation [3, 16]. The source emission pattern plays a double role since it influences not only the direct path but also the whole Room Impulse Response (RIR), controlling the amount of energy irradi- ated along each propagation path. Consequently, the RIR, which fully describes the point-to-point acoustic propagation in a given This work was partially supported by the European Commission under the Project SCENIC, Future Emerging Technologies (FET), 7th FP, grant number 226007. 1 Further details are available at: http://www.thescenicproject.eu enclosure, could be used to extrapolate the properties of the source emission pattern. An accurate derivation of the RIR from the cap- tured audio signal is possible only if the originally emitted sound is known. Alternatively blind channel identification methods are adopted [8], which typically operate using arrays of microphones. In this work instead we are interested in using the information pro- vided by the GCC-PHAT [9] computed at several surrounding mi- crophone pairs. It has been shown that such information is related to the orientation of a non omnidirectional source and that it can be employed to infer the direction of sound emission [4]. The radia- tion pattern influences in a very articulated way the behaviour of the GCC-PHAT, affecting not only the direct path but the whole multi- path propagation. Therefore, deriving a direct relationship between the emission pattern and the observed GCC-PHAT measurements is not trivial. In this work we suggest adopting an environment aware method, which, provided that accurate descriptors of the environ- ment are available, approximates the acoustic propagation by con- sidering low order microphone mirrors [2] and consequently models the corresponding expected GCC-PHAT. A similar approach was followed in [14] which presents a method, relying on first order mirror images, for source orientation estimation using eigenvalues of the cross-correlation matrix. The problem of estimating the radiation pattern of an acoustic source has not received much attention by the research community so far. Few works are available in the literature on the topic. In [11] the acoustic source is approximated as a circular piston whose ra- dius determines the emission pattern. Based on the energy received at a microphone array the piston radius is derived. More recently in [12] a rough representation of the radiation pattern, in combi- nation with the source position and orientation, is derived from a weighted delay-and-sum beamforming. This paper is organized as follows. Section 2 introduces the adopted acoustic propagation model in presence of a non omnidirec- tional source. Section 3 presents the maximum likelihood estima- tion framework whose performance, measured on simulated data, is discussed in Section 4. A final discussion concludes the paper. 2. ACOUSTIC PROPAGATION WITH DIRECTIVE SOURCES Let us assume that N microphones are available in a given enclosure where a directional acoustic source is active with a certain orienta- tion. Each microphone is identified by the couple (r n ,θ n ) indicating its distance r n and azimuth θ n with respect to the source (see Fig- ure 2). The azimuth corresponds to the angular distance between the direction the source is aimed at and the line connecting the source and the microphone. Assuming an FIR modeling of the RIR h n (t ) between the source and the n-th microphone, a common approach is to split the acoustic propagation into the direct path and the rever- berated part: h n (t )= h n,d δ (t - r n /c)+ h n,r (t ) (1) where the attenuation factor h n,d includes both the propagation loss and the directivity gain (in case the source is not omnidirectional), while c is the speed of sound. The reverberant part h n,r (t ) includes 19th European Signal Processing Conference (EUSIPCO 2011) Barcelona, Spain, August 29 - September 2, 2011 © EURASIP, 2011 - ISSN 2076-1465 151