Page  107 ï~~Real-Time Virtual Audio Reality Lauri Savioja 1, Jyri Huopaniemi 2, Tommi Huotilainen 2,3, and Tapio Takala 1 SHelsinkiUniversity of Technology Laboratory of Computer Science Otakaari 1, FIN-02150 Espoo, Finland 2Helsinki University of Technology 3ABB Industry, Pulp & Paper Acoustics Laboratory Otakaari 5 A, FIN-02150 Espoo, Fin- P.O. Box 94, FIN-00381 Helsinki, land Finland Lauri. Savioja@hut. fi, Jyri.Huopaniemi@hut. fi, fi, tta@cs.hut. fi Abstract A real-time virtual audio reality model has been created. The system includes model-based sound synthesizers, geometrical room acoustics modeling, binaural auralization for headphone or loudspeaker listening, and hiqh-quality animation. This paper discusses the following subsystems of the designed environment: The implementation of the audio processing soft- and hardware, and the design of a dedicated multiprocessor DSP hardware platform. The design goal of the overall project has been to create a virtual musical event that is authentic both in terms of audio and visual quality. Novelties of this system include a real-time image-source algorithm for rooms of arbitrary shape, shorter HRTF filter approximations for more efficient auralization, and a network-based distributed implementation of the audio processing soft- and hardware. 1 Introduction Application fields of virtual audio reality environments include computer music, acoustics, and multimedia. There are often computational constraints that lead to very simplified systems that faintly resemble the physical reality. We present a distributed expandable virtual audio reality system that can accurately yet efficiently model room acoustics and spatial hearing in real time. In Chapter 2, real-time binaural modeling of room acoustics is discussed. Digital signal processing (DSP) aspects of room acoustics and head-related transfer function (HRTF) implementation are overviewed in Chapter 3. In Chapter 4, the implementation of the system is described. 2 Real-time Binaural Room Acoustics Modeling Computers have been used nearly thirty years to model room acoustics [Krokstad, 1968]. A good overview of current modeling algorithms is presented in [Kuttruff, 1995]. Performance issues play an important role in the making of a real-time application [Kleiner et al., 1993] and therefore there are quite few alternative algorithms available. Methods that try to solve the wave equation are far too slow for real-time purposes. Ray tracing and image source methods are the most often used algorithms which base on the geometrical room acoustics. Of these, the image source method is faster for modeling low-order reflections. For auralization purposes simulations must be done binaurally. Some binaural simulation methods are presented in articles [Lehnert, 1992] and [Martin, 1993]. The image source method is good for binaural processing, since the incoming directions of sounds are the same as the orientations of the image sources. An example of a geometrical concert hall model is illustrated in Fig. 1. 2.1 The Image Source Method The image source method is a geometrical acoustics based method and is widely used to model room acoustics. The method is thoroughly explained in many articles [Allen et al., 1979], [Borish, 1984]. The algorithm implemented in this software is quite traditional. There are although some enhancements to achieve a better performance level. In the image source method the amount of image sources grows exponentially with the order of reflections. Therefore it is necessary to calculate only the image sources that might come visible during the first reflections. To achieve this we make a preprocessing run with ray tracing to check visibilities of all surface pairs. 2.2 Real-Time Communication In our application the real-time image source calculation module communicates with two other processes. It gets input from the graphical user interface. This input represents the movements of the listener. The model generates output for the auralization unit. To calculate the image sources the model needs following input information: 1) the geometry of the ICMC Proceedings 1996 107 Savioja et al.

Page  108 ï~~Figure1. The Sigyn Hall in Turku, Finland, is one of the halls where simulations were carried out. room, 2) the materials of the room, 3) the location of the sound source, and 4) the location and orientation of the listener. The model calculates positions and orientations of image sources. A following set of parameters concerning each image source is passed to sound processor: 1) the distance from the listener, 2) the azimuth angle to the listener, 3) the elevation angle to the listener, and 4) two filter coefficients which describe the material properties in reflections. The amount of image sources depends on the available computing capacity. In our real-time solution typically 20-30 image sources are passed forward. The model keeps track of the previously calculated situation. Newly arrived input is checked against that. If changes in any variable are large enough, some updating process is necessary. 2.3 Updating the Image Sources The main principle in the updating process is that the system must respond immediately to any changes in the environment. That is reached by gradually refining calculation. In the first place only the direct sound is calculated and its parameters are passed to the auralization process. If there are no other changes queuing to be processed first order reflections are calculated, and then second order, and so on. In a changing environment there are three different possibilities that may cause recalculations. Movement of the sound source If the sound source moves, all image sources must be recalculated. The same applies also to the situation when something in the environment, such as a wall, moves. Movement of the listener The visibilities of all image sources must be validated whenever the listener moves. The locations of the image sources do not vary and therefore there is no need for recalculation. Turning of the listener If the listener turns without moving there are no changes in the positions of the image sources. Only the azimuth and elevation angles may change and those must be recalculated. 2.4 Material Parameters Each surface of the modeled room has been given sound absorption characteristics, generally in octave bands from 125 Hz to 4000 Hz. In a real-time implementation, these frequency-dependent absorption characteristics are taken into account by designing first-order IIR approximations to fit the magnitude response data of each reflection coefficient combination. 3 Auralization Issues The goal in real-time auralization is to preserve the acoustical characteristics of the modeled space to such extent that the computational requirements are still met. This places constraints to the accuracy and quality of the final auditory illusion. The steps required in our auralization strategy can be divided in the following way: 1) model the first room reflections with an image-source model of the concert hall, 2) use accurate HRTF processing for the direct sound, 3) apply simplified directional filtering for the first reflections, and 4) create a recursive reverberation filter to model late reverberation. 3.1 Real-Time Room Impulse Response Processing The use of methods based on geometrical room acoustics in real-time modeling of the full room impulse response is out of the calculation capacity of modern computers. To solve this problem, hybrid systems that exhibit the same behavior as room impulse responses in a computationally efficient manner have to be found. We ended up using a recursive digital filter structure based on earlier reverberator designs [Schroeder, 1962] [Moorer, 1979], which is computationally realizable yet gains good results [Huopaniemi et al., 1994]. The structure combines the implemented image-source method and late reverberation generation. The early reverberation filter is a tapped delay line with lowpass filtered outputs designed to fit the early reflection data of a real concert hall. The recursive late reverberation filter structure is based on comb and alipass filters. 3.2 HRTF Filter Design Sound source localization is achieved in a static case primarily with three cues [Blauert, 1983]: 1) the interaural time difference (ITD), 2) the interaural amplitude difference (LAD), and 3) the frequencydependent filtering due to the pinnae, the head, and the torso of the listener. The head-related transfer function (HRTF) represents a free-field transfer func tion from a fixed point in a space to a point in the test person's ear canal. There are often computational Savioja et al. 108 ICMC Proceedings 1996

Page  109 ï~~constraints that lead to the need of HRTF impulse response approximation. This can be carried out using conventional digital filter design techniques. In most cases, the measured HRTFs have to be preprocessed in order to account for the effects of the loudspeaker, microphone, (and headphones for binaural reproduction) that were used in the measurement. Further equalization may be applied in order to obtain a generalized set of filters. Such equalization methods are free-field equalization and diffuse-field equalization. Smoothing of the responses may also be applied before the filter design. A method called cepstral smoothing has been used [Huopaniemi et al., 1995], which is implemented by proper windowing of the real cepstrum. This smoothing method may also yield minimum-phase results. Minimum-phase Reconstruction An attractive solution for HRTF modeling is to reconstruct data-reduced minimum-phase versions of the modeled HRTF impulse responses. A mixedphase impulse response can be turned into minimum-phase form without affecting the amplitude response. The attractions of minimum-phase systems in binaural simulation are: 1) the filter lengths are the shortest possible for a specific amplitude response, 2) the filter implementation structure is simple, 3) minimum-phase filters perform better in dynamic interpolation. According to Kistler and Wightman [1993], minimum-phase reconstruction does not have any perceptual consequences. With minimum-phase reconstructed HRTFs, it is possible to separate and estimate the ITD of the filter pair, and insert the delay as a separate delay line to one of the filters in the simulation stage. The delay error due to rounding of the ITD to the nearest unitdelay multiple can be avoided using fractional delay filtering (see [Laakso et al., 1996] for a comprehensive review on this subject). FIR and IIR Filter Implementations Digital filter approximations (FIR and IIR) of HRTFs have been studied to some extent in the literature over the past decade. Filter design using auditory criteria (which is desired because we are interseted in audible results) have been proposed by quite few authors, however. There are two alternatives to a non-linear frequency scale approach: 1) weighting of the error criteria, and 2) frequency warping. In the latter, warping is accomplished by resampling the magnitude spectrum on a warped frequency scale [Smith, 1983] [Jot et al., 1995]. In practice, warping is implemented by bilinear conformal mapping with 1 st-order ailpass filters. The resulting filters have considerably better low-frequency resolution with a trade-off on high-frequency accuracy, which is tolerable according to the psychoacoustic theory. Filters may be implemented in the warped domain, or in a normal way by dewarping [Jot et al., 1995]. 103 Frequency / Hz Figure 2. HRTF modeling (source: Briiel&Kjaer BK4100 dummy head, right ear, 0Â~ elev, 300 azim). Solid line: original, dashed line: 90-tap FIR, dotted line: IR order 44, dash-dot line: WIIR order 12. A similar non-uniform frequency resolution approach has been taken in the HRTF modeling research at the HUT Acoustics Laboratory. Warped FIR (WFIR) and IIR (WIIR) structures have been applied to HRTF approximation [Huopaniemi and Karjalainen, 1996]. Results of example HRTF filter design are presented in Figure 2. The 90-tap FIR filter has been designed using a rectangular window. IIR and WIIR designs were carried out using the time-domain Prony's method. The magnitude scale is relative so that differences can be illustrated. It can be clearly seen that the warped structure retains the essential features of the magnitude response even at low filter orders (order 12 in this case). 3.3 Real-Time Auralization The auralization system obtains the following input parameters which are fed into the computation: " direct sound and image-source parameters " HRTF data for the direct sound and directional filters (minimum-phase WFIR or WIIR implementation stored at 10Â~ azimuth and elevation intervals) " "dry" audio input from a physical model or an external audio source The output of the auralization unit is at present directed to headphone listening (diffuse-field equalized headphone, e.g., AKG K240DF), but software for conversion to transaural or multispeaker format has also been implemented. 4 System Implementation We have used a distributed implementation on an Ethernet-based network to gain better computational power and flexibility. Currently we use one Silicon Graphics workstation for real-time visualization and the graphical user interface (GUI) and another for image-source calculations. We have also used a Texas Instruments TMS320C40-based signal processor system that performs direction- and frequency ICMC Proceedings 1996 109 Savioja et al.

Page  110 ï~~Silicon Graphics for GUI Silicon Graphics for Image Source Computation AT/ Ethrae Figure 3. The distributed implementation over an Ethernet-based network dependent filtering and ITD for each image source, the recursive reverberation filtering, and the HRTF processing. The basic idea for the Ethernet-based system is to use the multiprocessor system as a remote controlled signal processing system. In the transfer process, the audio source signal and/or control parameter block is transmitted through the network to the signal processing system, which receives the data, processes the audio signal and sends the stereophonic audio result back to the workstation in real time (Fig. 3). 5 Summary We have developed a soft- and hardware system for producing virtual audiovisual performances in realtime. The listener can freely move in the virtual concert hall where a virtual musician plays a virtual instrument. Early reflections in the concert hall are computed binaurally with image-source method. For late reverberation we use a recursive filter structure consisting of comb and allpass filters. Auralization is done by using the interaural time difference (ITD) and head-related transfer functions (HRTF). References [Allen et al., 1979] J. Allen and D. Berkley. Image method for efficiently simulating small-room acoustics. J. Acoust. Soc. Am., 65 (4), pp. 943-950, April 1979. [Borish, 1984] J. Borish. Extension of the image model to arbitrary polyhedra. J. Acoust. Soc. Am., 75 (6), pp. 1827-1836, 1984. [Blauert, 1983] J. Blauert. Spatial Hearing. M.I.T. Press, Cambridge, MA, 1983. [Huopaniemi et al., 1994] J. Huopaniemi, M. Karjalainen, V. Valimaki, T. Huotilainen. Virtual instruments in virtual rooms - a real-time binaural room simulation environment for physical models of musical instruments. Proc. 1994 Int. Computer Music Conf., pp. 455-462, Arhus, Denmark, 1994. [Huopaniemi et al., 1995] J. Huopaniemi, M. Karjalainen, V. Vailimaki. Physical models of musical instruments in real-time binaural room simulation. In Proc. Int. Congr. on Acoustics (ICA '95), vol. III, pp. 447-450, Trondheim, Norway, 1995. [Huopaniemi and Karjalainen, 1996] J. Huopaniemi and M. Karjalainen. HRTF filter design based on auditory criteria. To be published in: Proc. Nordic Acoustical Meeting (NAM'96), Helsinki, 1996. [Jot et al., 1995] J.-M. Jot, V. Larcher, and 0. Warusfel. Digital signal processing issues in the context of binaural and transaural stereophony. Presented at the 98th AES Conv., preprint 3980 (E-2), Paris, France, 1995. [Kistler and Wightman, 1992] D. Kistler and F. Wightman and. A model of head-related transfer functions based on principal components analysis and minimum-phase reconstruction. J. Acoust. Soc. Am., 9 1 (3), pp. 1637-1647, 1992. [Kleiner et al., 1993] M. Kleiner, B.-I. Dalenback, and P. Svensson. Auralization - an overview. J. Audio Eng. Soc., 41(11), pp. 861-875, 1993. [Krokstad, 1968] A. Krokstad, S. Strom, and S. Sorsdal. Calculating the acoustical room response by the use of a ray tracing technique. J. Sound Vib., 8 (1), pp. 118-125, 1968. [Kuttruff, 1995] H. Kuttruff. Sound field prediction in rooms. In. Proc. Int. Congr. on Acoustics (ICA '95), pp. 545-552, 1995. [Laakso et al., 1996] T. I. Laakso, V. Valimaki, M. Karjalainen, and U. K. Laine. Splitting the unit delay - tools for fractional delay filter design. IEEE Signal Processing Magazine, 13 (1), pp. 30-60, 1996. [Lehnert, 1992] H. Lehnert and J. Blauert. Principles of Binaural Room Simulation. Applied Acoustics, 3 6 (3-4), pp. 259-291, 1992. [Martin, 1993] J. Martin, D. Van Maercke, and J.-P. Vian. Binaural simulation of concert halls: a new approach for the binaural reverberation process. J. Acoust. Soc. Am., 94 (6), pp. 3255-3264, 1993. [Moorer, 1979] James A. Moorer. About this reverberation business. Computer Music J., 3 (2), pp. 13 -28, 1979. [Schroeder, 1962] M. Schroeder. Natural sounding artificial reverberation. 1. Acoust. Soc. Anm., 10 (3), 1962. [Smith, 1983] J. Smith, Techniques for digital filter design and system identification with application to the violin, Ph.D. dissertation, CCRMA, Department of Music, Stanford University (Standord, CA, 1983). Savioja et al. 110 ICMC Proceedings 1996