SuRIR: A Dataset of Successively Measured Room Impulse Responses for Localizing Silent Humans
Year: 2026
Type: dataset
Abstract: Acoustic localization of humans traditionally focuses on scenarios in which they emit sound. In certain settings, however, a human may remain silent while another sound source is active. Recent work demonstrated that silent humans can be localized by analyzing the subtle perturbations their bodies introduce into room impulse responses (RIRs). The existing SoundCam dataset includes RIRs estimated with a silent human present, but provides only single-snapshot RIRs per configuration. In contrast, successively measured RIR estimates enable the analysis of subtle temporal variations caused by involuntary human micromotion, e.g., breathing, which can lead to more robust localization performance than single-snapshot human‑present vs. human-absent comparisons. To advance research in this field, we introduce SuRIR, the first dataset that provides four successive RIR estimates for each configuration, measured a few seconds apart. SuRIR comprises 14,400 RIR estimates, acquired at 48 kHz using exponential sine sweeps. Measurements were conducted using six microphones and two loudspeakers across two rooms, with three participants performing five distinct movement patterns. To provide a static reference condition, SuRIR additionally includes 960 RIRs collected using a mannequin. Besides silent human localization, SuRIR provides a controlled testbed for evaluating RIR-based algorithms under subtle, real-world environmental variations. The associated paper can be found here. Dataset Structure The dataset contains successively measured room impulse responses (RIRs) for acoustic localization of silent humans. The data are provided in the SOFA format using the SingleRoomSRIR convention. The dataset comprises 14,400 RIR estimates of 1 s duration from two rooms, three human participants, five human positions, five experimental conditions, two loudspeakers, six microphones, four repetitions, and two sweep lengths (1 s and 5 s). In addition to the human measurements, 960 RIRs using a motionless mannequin are provided as a reference condition. Human measurements The human measurements are organized into individual SOFA files according to room, participant, experimental condition, and sweep length. The filenames follow the structure RoomA_Participant01_Condition01_Sweep1s.sofa where RoomA identifies the room, Participant01 identifies the participant, Condition01 identifies the experimental condition, and Sweep1s identifies the sweep duration used for RIR estimation. The two sweep lengths are stored in separate SOFA files, allowing measurements obtained with the different sweep durations to be evaluated independently. The five experimental conditions vary in terms of the participant's posture, movement, and orientation. Specifically, they are: Standing while moving, facing left Standing still, facing left Standing still, facing down Sitting while moving, facing left Sitting still, facing left Here, moving denotes subtle whole-body sway and facing refers to the orientation of the participant's entire body rather than that of the head alone. Left indicates facing the left wall, while down denotes facing parallel to the left wall toward the bottom of the floor plans.The conditions are shown at times 9:49-10:04 in the following video. In total, there are 60 SOFA files containing human measurements: 2 rooms × 3 participants × 5 conditions × 2 sweep lengths = 60 files Each human-measurement SOFA file contains 40 measurements with six RIR channels each, i.e., 40 × 6 individual RIRs. The 40 measurements result from the combination of: 5 human positions, 2 loudspeakers, and 4 repetitions per loudspeaker and human position. The measurements are ordered first by human position, then by loudspeaker, and finally by repetition. For each human position, the first four measurements correspond to the four repetitions (Rx) recorded with loudspeaker 1 (LS1), followed by the four repetitions recorded with loudspeaker 2 (LS2). Thus, the measurement sequence is LS1-R1, LS1-R2, LS1-R3, LS1-R4, LS2-R1, LS2-R2, LS2-R3, LS2-R4, followed by the same sequence for human positions 2–5. The repetitions are intentionally retained as individual measurements and are not averaged. This preserves the small changes between repeated RIRs that may arise from human micromotion and are relevant to the intended localization task. Mannequin reference measurements In addition to the human recordings, the dataset contains measurements obtained with a motionless mannequin. These recordings provide a reference condition without micromotion and can therefore be used to investigate changes in the RIRs that are unrelated to human movement. The mannequin measurements are provided in four SOFA files following the naming structure RoomA_Mannequin_Sweep1s.sofa where the room and sweep duration are encoded in the filename. Each mannequin SOFA file contains 40 measurements with six RIR channels each, following the same ordering as the human-measurement files: five human positions, two loudspeakers, and four repetitions. Rooms The dataset includes .pdf floor plans for both rooms, along with the loudspeaker, microphone, and participant positions. Room A corresponds to a furnished living room, whereas Room B is an unfurnished space. Further details are provided in the accompanying Forum Acusticum 2026 paper. SOFA metadata All SOFA files contain spatial and experimental metadata associated with the measurements. This includes the positions and orientations of the human, microphones, and loudspeakers, as well as information describing the room geometry. Spatial coordinates are specified in a Cartesian coordinate system in metres. The room geometry is represented by the four corner coordinates of the non-rectangular room floor plan together with the room height. Additional metadata describe the room volume and temperature and identify the dataset, measurement setup, and relevant references. The RIRs are stored as finite impulse responses (FIRs) with a sampling rate of 48 kHz. The SOFA files were generated using the sofar Python package. Dataset summary The complete dataset consists of 64 SOFA files containing 15,360 individual RIRs. The dataset is structured such that each SOFA file represents a fixed room, human identity (human participant or mannequin), experimental condition, and sweep duration, while human position, loudspeaker, and repetition vary within the file.
Language: English
Topic: Speech and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7208770872