Where to Put That Microphone? A Study of Sound Localization in Ambisonics Recordings

Where to Put That Microphone? A Study of Sound Localization in Ambisonics Recordings Maham Riaz1,JinyueGuo 1, Ça!rı Erdem2, and Alexander Refsum Jensenius 1 1RITMO Centre, Department of Musicology, University of Oslo, Norway 2Department of Informatics, University of Oslo, Norway {mahamr,jinyueg,cagrie,alexanje}@uio.no Abstract. This paper examines the e!ects of microphone placement on sound localization in first-order Ambisonics recordings. Two microphone setups were used to capture a moving audio source in a lab environment. Array A, a tetrahedral microphone, was placed in the centre of the recording space. Array B consisted of four similar tetrahedral microphones charting a rectangular perimeter surrounding the space. Motion capture data of the moving sound source shows that anglegrams calculated from the Ambisonics recordings can be e!ectively used for sound localization. An additional perceptual listening study with binaural renders of the audio signals showed that the centrally-placed Array Aprovidedsuperiorlocalization.However,thecorner-placedArrayB performed better than expected. Keywords: Ambisonics ·Motion capture ·Spatial audio. 1Introduction Fig. 1. The Zoom H3-VR, a firstorder Ambisonics recorder, with a motion capture marker on top. Our ongoing research concerns recording and analysing public and private indoor soundscapes. Ambisonics recorders provide a reliable solution for capturing soundscapes and serving as source material for room characterization and sound localization. However, a recurring question is where the recorders should be placed in the space. The general rule of thumb is to set the microphone in the middle of the room. However, that may not be feasible due to a room’s construction, objects in the way, or people’s position and motion. All rights remain with the authors under the Creative Commons Attribution 4.0 International License (CC BY 4.0). Proc. of the 17th Int. Symposium on Computer Music Multidisciplinary Research, London, United Kingdom, 2025 Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 455 M. Riaz et al. This research lies at the intersection of music technology and psychology, combining audio processing with human perception to explore acoustic environments. The paper reports on an experiment integrating principles from sound recording, digital signal processing, and auditory perception. Traditional algorithms and methods for sound localization often fail in real-world, non-studio situations. Instead of identifying a single optimal listening position, this study examines recordings of a moving sound source from di"erent locations within the same environment aiming to answer the following questions: RQ1 To what extent does microphone placement matter when localizing a moving audio source with Ambisonics? RQ2 Does the combination of anglegrams with motion capture data work well for sound localization comparisons? RQ3 What are the localization di"erences between microphones placed in the corners of a room versus the centre? The following sections review relevant literature, followed by a description of our lab experiment and listening study. 2Background 2.1 Soundscape Studies Soundscape studies investigate how humans perceive and interact with their auditory environments [18]. While early research focused on outdoor soundscapes— e.g., comparing urban noise to natural soundscapes—recent research increasingly recognizes the psychological and social e"ects of indoor soundscapes [13]. Outdoor soundscapes typically comprise sources such as tra#c, weather, and wildlife, whereas indoor soundscapes tend to feature more localized and controlled sound sources, including speech, ventilation systems, and reverberant e"ects. Soundscapes have also been used as aesthetic constructs by composers and sound artists who treat environmental sound as compositional material. Schafer used outdoor environments as performative spaces in his experimental opera The Princess of the Stars [19]. Electroacoustic Works by Hildegard Westerkamp transformed urban field recordings into sound narratives, exploring how amplification or attenuation of selective sonic details can reshape a space’s emotional feel [1]. Similarly, Barry Truax theorised the practice of “soundscape composition,” arguing that studio-based manipulations of everyday acoustic environments can reveal musical structures in what might otherwise be dismissed as noise [20]. 2.2 Foreground vs Background Classifying sound sources into foreground and background elements is central to what is often referred to as “auditory scene analysis" [4]. Foreground sounds (such as a mobile phone ringing) demand immediate attention, while background Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 456 Sound Localization in Ambisonics Recordings sounds (such as the hum of a ventilation system) form a continuous, often subconscious sonic backdrop. This distinction is critical in applications ranging from urban noise mitigation to architectural acoustics [6], yet it raises the question of how to e"ectively record and analyse complex soundscapes. The literature on techniques for recording indoor soundscapes, which acknowledges the distinction between foreground and background elements and their transitional relationships, is sparse. While studies exist—often in outdoor environments [18] or controlled indoor lab measurements—few address realworld, unstructured settings [13]. This gap contrasts sharply with music production practices, where ambient room microphones are used to capture background spatial reflections [12] and spot microphones are used for foreground musical instruments. Similarly, game audio utilizes Ambisonics for 360-degree background sound and audio objects for foreground elements. 2.3 Room Acoustics The acoustics of confined spaces amplify and mask di"erent frequencies [3], often referred to as spectral “coloration.” No two locations within a room share identical physical properties, as even minor variations in materiality and geometry create unique perceptual outcomes. In practice, perfectly uniform indoor environments do not exist. The variability is quantified through the room impulse response (RIR), which maps how aspace“colours”animpulsesignal.RIRscontaindirectsound,earlyreflections, and late reverberation. Another standard metric is the RT60, the time required for reverberant energy to decay by 60 dB [8]. Yet another metric is the directto-reverberant energy ratio: this is the ratio of signal energy to the reverberant energy present in the room as a result of that signal [14]. Architecture and construction materials strategically shape the acoustic character in spaces designed with acoustics in mind, such as cathedrals and concert halls. High-, mid-, and low-frequency di"users or absorbers are implemented to manage sound reflections, influencing factors like reverberation and decay time. Additional elements such as seating, audience presence, carpeting, and curtains a"ect sound absorption and dispersion [7]. The varying surface absorption coefficients contribute to the distinctive acoustic characteristics of a space. Sound waves interact with surfaces and structures, altering their propagation and creating an acoustic “fingerprint.” This interaction changes the perception of sound at di"erent locations within the same space: the early reflection paths are reconfigured, and the direct-to-reverberant energy ratio also changes accordingly. 2.4 Ambisonics Ambisonics is a comprehensive spatial audio technology that encompasses microphone design, audio encoding, and playback for capturing an immersive 360degree sound field [11]. The recorded audio can be rendered binaurally for headphone playback or in surround sound configurations, such as 5.1, 7.1, and higher. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 457 M. Riaz et al. As the “order” of Ambisonics increases, so does the channel count, enhancing spatial resolution. The Ambisonics order is calculated using the following formula: (N+ 1)2. Thus, the simplest form (first-order) uses four channels. Ambisonics recording devices consist of multiple cardioid microphone capsules, with the total number depending on the Ambisonics order. The raw directional recordings are in “A-format,” pending subsequent encoding into the most common “B-format.” The encoding process uses normalization and weighting algorithms to incorporate all channels, capturing detailed spatial information. The two prevalent encoding schemes are AmbiX and FuMa, which di"er in their channel orders and normalization weightings. B-format audio contains four channels: WXYZ. The W channel stores omnidirectional sound, while the XYZ represent directional sound on the front–back, left–right, and up–down axes, respectively. If the A-format recordings are encoded to B-format with the AmbiX channel format, the resultant signal follows the WYZX channel order [22]. 2.5 Localization Comparisons While Ambisonics is not a new topic in music technology research, it is rarely combined with motion capture for sound localization studies. This study introduces a novel method that utilizes an infrared mocap system to record the precise location of the moving audio source, serving as the “ground truth.” This crossmodal comparison between audio signal localization (through anglegrams) [17] and actual spatial location (through mocap) aims to investigate the perception of sound location versus its exact position. Previous sound localization studies have combined LED-based mocap with binaural audio recording [9]. In contrast, our study employs infrared motion capture in a laboratory equipped with ceiling-mounted cameras focused on a calibrated space. Spherical reflective markers attached to the tracked objects capture their position in space using Cartesian coordinates. Infrared flashes emitted by the cameras during the motion capture session are recorded after reflecting o" the markers, allowing precise tracking of the objects’ motion. Human sound localization heavily relies on binaural cues such as the interaural time di"erence (ITD) and the interaural intensity di"erence (IID) between the two ears [16]. These cues are preserved in binaural renders of Ambisonic audio, enabling listeners to perceive the direction and motion of the sound source. However, certain regions around the human head (or, in this case, the microphone) can introduce uncertainty in localization perception, known as cones of confusion [2]. These occur when two sound sources equidistant from the listener have similar ITDs and IIDs, leading to perceptual ambiguities in localization. Head movement mitigates this ambiguity in real-life scenarios [21]. Several studies on binaural sound localization have examined the average degree of error in human listeners. One study, incorporating head motion, reported a 15.4-degree error [5], while another calculated a maximum of 20 degrees [15]. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 458 Sound Localization in Ambisonics Recordings 3 Data Collection The study reported here integrates multi-microphone Ambisonics recordings with infrared mocap, providing a “cross-modal” approach to examining how recorded signals are perceived in di"erent spatial conditions. Three motion capture sessions, each corresponding to a di"erent height level, were recorded, each accompanied by a video recording. This section outlines the methods used. 3.1 Microphone Setup and Recording 5.0 m Mic A Mic B2Mic B1 Mic B3 Mic B4 4.0 m radius = 1.60 m Front of the Lab x axis of mocap y axis of mocap Fig. 2. An aerial view of the lab with microphones located in the centre (Array A) and corners (Array B). The circular trajectory of the moving audio source is sketched. The experimental setup featured two microphone arrays to capture the moving audio source. We used a total of five Zoom H3-VR recording devices, each equipped with four microphone capsules arranged in a tetrahedral pattern to achieve first-order Ambisonics (Figure 1). A sketch of the setup is shown in Figure 2, with Array A consisting of one recorder located in the centre. Array B, positioned at each of the four corners of the space, consisted of four recorders. As a result, Array Ahad4channels,andArrayBhad 4→4 = 16 channels. All microphones were oriented forward, facing the front of the lab (RT60 of 0.3 seconds [17]). The microphone capsules of the Array A recorder were roughly aligned directionally with one capsule from each of the four microphones in Array B. The microphones were set to a gain level of 90 (unitless) and recorded at a sample rate of 48 kHz with 24 bit resolution. The audio was initially recorded in A-format Ambisonics and converted to B-format AmbiX with the standalone Zoom Ambisonics Player. 3.2 Moving Audio Source Since the objective of the experiment was sound localization, we decided to use a chirp sound played in a loop at constant amplitude (Figure 3). The sound was played from a mobile phone carried by one of the experimenters as they moved in a circular path in the recording space. Two complete circles were performed at three heights: Height (170 cm), Height/2(85 cm), and Height↑1.5(255 cm). The heights were chosen based on the average human height of 170 cm [10]. Moving in a circle ensures that azimuth changes are recorded in the signal. Additionally, using di"erent heights helps expand the evaluation of the recording Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 459 M. Riaz et al. Fig. 3. Spectrogram of the chirp signal as recorded by microphone A at Height level. methods to include di"erent elevation levels. The approximate radius of the motion trajectory was 160 cm, which was the midpoint of the straight-line distance between the centre and the corner of the floor space (Figure 2). 3.3 Motion Capture Setup and Recording Marker-based, infrared mocap was used to track the position of the moving audio source (“ground truth”). This was accomplished using a Qualisys system with 11 Oqus 300/310 cameras mounted at ceiling height, covering a calibrated floor area of approximately 5m by 4m.Eachofthefivemicrophoneshadareflective marker (Figure 1), while the moving audio source had three, allowing tracking of its position, orientation, and rotation. The motion capture system’s sampling frequency was set to 250 Hz. Fig. 4. The trace of the moving audio source at Height level. After cleaning and gap-filling the motion capture data, the sessions were trimmed to the desired duration for analysis. The software automatically trimmed the accompanying video files, facilitating the synchronization process. The exported .tsv motion capture files were used for localization analysis in Python. Figure 4 shows the trace of the moving audio source from a motion capture session. 3.4 Syncing the Files Each motion capture session includes a complimentary video recording with an audio track. These audio tracks Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 460 Sound Localization in Ambisonics Recordings were aligned with the recordings from Array A and B in Reaper and Audacity using peak matching to claps performed at the beginning of the recording. The audio files for each experimental procedure were exported for further analysis in Python. 3.5 Listening Test Two expert listeners (with a professional background in music production) and two non-experts were recruited for a perceptual listening task. For this test, signals from Array A (centre) and Array B (corners) across the three height levels were rendered into binaural and stereo .wav files for perceptual listening assessments using the Zoom Ambisonics Player. The signals were also downmixed to mono in Reaper as .wav files. Playback was conducted through headphones to ensure clear spatial resolution. 4Results The audio signals were subsequently decoded to generate anglegrams, graphical plots that represent the azimuth and elevation changes over time of a moving audio source. These plots were visualized in Python for comparison between audio recordings and the mocap data. 4.1 Anglegrams In this study, we analyzed the spatial information of the signals using “anglegrams,” a 2D graphical representation that illustrates the azimuth and elevation changes of a moving audio source relative to the microphone (Figure 5). The anglegrams are calculated as follows: 1. Divide the space around the microphone into a grid of evenly spaced points. Azimuth angles are divided every 2→from ↓180→to 178→ (e.g., ↓180→,↓178→,↓176→,...,178→). The elevation angles consist of every 2→from ↓90→to 90→(e.g., ↓90→,↓88→,...,90→). This yields a total of 180 →90 = 16,200 possible points. 2. Decode Ambisonics for each virtual speaker. For each possible location of the virtual speakers, we decode the Ambisonics signal at the targeted direction. This creates 16,200 mono audio tracks, one for each direction. 3. Measure energy in short time windows. Split the audio into short overlapping frames. For each frame, calculate the RMS energy, which yields the energy map E(n, ω,ε), where ndenotes the time frame, ωthe azimuth angle, and εthe elevation angle. 4. Find the loudest direction per frame. For each time frame, compare the energy values across all 16,200 directions. The direction with the highest energy is labeled as the dominant sound source for that instant: A(n)= (ωn,ε n,E n)=(argmax ω,εE(n, ω,ε),max ω,εE(n, ω,ε)).We define the resulting sequence A(n)as the anglegram. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 461 M. Riaz et al. A↓35 dB threshold was applied when generating the anglegrams to mitigate noise originating from the experimenter’s light footsteps. 4.2 Sound Localization in Array A Versus MoCap The anglegram for Array A (in the centre) at Height level is illustrated in Figure 5. The localization values for azimuth and elevation closely match those of the motion capture recording, with an average error of 0.22 rad (12.6→)and 0.13 rad (7.4→), respectively. The potential for mislocalization increases when the height decreases or increases. For a comprehensive overview, please refer to Supplementary Material. Fig. 5. The anglegram of Array A showing the error di!erence against the motion capture values, at Height level. The color of the points in the graph denotes di!erent amplitudes. The pink data points, manifesting as straight lines in the graph, are from the motion capture values of the physical location of the moving audio source. The orange points depict the error di!erences between the angle calculations from the microphone audio and the motion capture ground truth. 4.3 Sound Localization in Array B Versus MoCap The analysis of Array B (corner microphones) focused on localizing a signal from various positions within the room. Table 1 shows the mean errors for both azimuth and elevation between the signals from Array B and the motion capture ground truth. Overall, the microphones do relatively better at Height/2 and Height levels for both azimuth and elevation. At Height level, the average Array B azimuth and elevation errors are 0.48 rad (27.5→)and0.31 rad (17.8→). Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 462 Sound Localization in Ambisonics Recordings 4.4 Array A Versus Array B As depicted in Table 1, Array A generally performs better than Array B, albeit with some exceptions. Microphone B2 exhibits lower localization errors than microphone A at Height/2for both azimuth and elevation. Similarly, microphone B4 yields better azimuth localization accuracy than microphone A at the Height ↑1.5level. Mean Azimuth Error (radians) Mean Elevation Error (radians) Microphone Height/2 Height Height*1.5 Height/2 Height Height*1.5 Mic B1 0.48 0.72 1.34 0.35 0.39 0.51 Mic B2 0.38 0.32 0.52 0.21 0.28 0.32 Mic B3 0.49 0.43 0.56 0.27 0.32 0.39 Mic B4 0.53 0.45 0.41 0.32 0.25 0.28 BMean 0.47 0.48 0.71 0.29 0.31 0.38 AMean 0.45 0.22 0.46 0.21 0.13 0.19 Table 1. Mean azimuth and elevation errors for each of the four microphones in Array B(corners)comparedwiththemeanofArrayBandA(centre). At the Height level, Array A’s azimuth and elevation errors measure 0.22 rad (12.6→)and0.13 rad (7.4→), respectively. In contrast, the mean error in Array B exceeds twice that amount. 4.5 Listening Results AcomprehensiveoverviewofthelisteningtestresultscanbefoundinSupplementary Material. Individual di"erences and varying expertise between the test participants resulted in variations in how spatial audio properties were perceived. Both expert listeners could hear the reflections of the chirp signal in binaural and stereo renders, but only one detected them in mono signals. Both experts could trace the trajectory of the moving audio source in binaural and stereo renders, mainly in azimuth changes, rather than elevation. Distance changes were noticeable in all signals. For non-experts, the reflections were less perceptible. Tracing the moving source’s trajectory yielded mixed results, with binaural and stereo signals performing better than mono. This was mostly true for azimuth changes, while elevation changes received inconsistent responses. Distance changes were generally easy to perceive. For all listeners, there was a general sense of direction and motion perceivable from the corner microphones. Microphones B1 and B4 predominantly channeled the signal to the right ear of the headphones. In contrast, microphones B2 and B3 favored the left ear, reflecting the position of the moving audio source relative to the microphones. Moreover, the sensation of rotation in trajectory and varying distance was perceptible due to fluctuations in loudness levels. Tracing the trajectory of the moving audio source was easier with microphone A positioned at the centre. Furthermore, the di"erences in height levels were perceived di"erently across all recordings, regardless of the microphone used. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 463