[Paper Review] MIRaGe: Multichannel Database Of Room Impulse Responses Measured On High-Resolution Cube-Shaped Grid In Multiple Acoustic Conditions

Jaroslav Čmejla, Tomáš Kounovský+3 more|arXiv (Cornell University)|Jul 29, 2019

Speech and Audio Processing15 references4 citations

TL;DR

MIRaGe is a high-resolution, multichannel database of room impulse responses (RIRs) measured on a dense 3D grid (4104 positions) within a 46×36×32 cm volume in a real acoustic lab, with three adjustable reverberation times. The database enables detailed analysis of beamforming robustness and spatial processing, demonstrating that real RIRs differ significantly from image-method simulations, especially in reverberation tails, and supports superior performance of supervised RTF identification methods like manifold projection in low-SNR conditions.

ABSTRACT

We introduce a database of multi-channel recordings performed in an acoustic lab with adjustable reverberation time. The recordings provide information about room impulse responses (RIR) for various positions of a loudspeaker. In particular, the main positions correspond to 4104 vertices of a cube-shaped dense grid within a 46x36x32 cm volume. The database thus provides a tool for detailed analyses of beampatterns of spatial processing methods as well as for training and testing of mathematical models of the acoustic field.

Motivation & Objective

• To address the lack of dense, real-world RIR measurements for evaluating spatial signal processing robustness under small source position perturbations.
• To provide a high-resolution, multichannel RIR database for training and testing mathematical models of acoustic fields.
• To enable detailed analysis of beampatterns and RTF variations across a 3D grid of source positions in a controlled, real acoustic environment.
• To offer a realistic alternative to image-method-based RIR simulators, which fail to capture complex reverberation characteristics of real rooms.
• To support the evaluation of advanced signal processing techniques, such as supervised RTF identification, under realistic acoustic conditions.

Proposed method

• A loudspeaker was precisely positioned at 4104 vertices of a cube-shaped grid (46×36×32 cm) with 2 cm spacing in x and y, and 4 cm in z.
• Six linear microphone arrays were placed at 1, 2, and 3 meters in front of the grid, and at -45° angles at the same distances, all at 115 cm height.
• RIRs were recorded for each source position using a 31st microphone mounted 2 cm in front of the loudspeaker, with the source directed toward the room center.
• Three reverberation time levels (T60 ≈ 100 ms, 300 ms, 600 ms) were achieved using adjustable acoustic absorption panels.
• Acoustic transfer functions (ATFs) and relative transfer functions (RTFs) were computed from the recorded signals using inverse filtering and least-squares estimation.
• The database was validated through two experiments: (1) beamforming blocking ability comparison between real and simulated RIRs, and (2) evaluation of supervised RTF identification (MP and NN) using real and simulated data.

Experimental results

Research questions

• RQ1 How do small changes in source position affect the spatial response and beamforming performance in real rooms?
• RQ2 To what extent do image-method-based RIR simulators accurately replicate the reverberation characteristics of real rooms?
• RQ3 Can supervised RTF identification methods like manifold projection outperform unsupervised methods in realistic, noisy, low-SNR conditions?
• RQ4 How does reverberation time influence the robustness of beamforming and RTF estimation in real acoustic environments?
• RQ5 What is the impact of RIR variability across a dense 3D grid on the performance of spatial signal processing algorithms?

Key findings

• The blocking ability of RTFs is higher when the reference microphone position is closer to the source position, confirming spatial sensitivity in beamforming performance.
• The overall blocking ability is lower at higher reverberation times (600 ms) due to the limited time-domain length of the RTFs (1536 taps), which affects temporal resolution.
• The real RIRs from the MIRaGe database exhibit significantly different reverberation tails compared to those generated by the image method, indicating that simulated RIRs cannot fully replicate real acoustic behavior.
• In low-SNR conditions (below 0 dB), both Nearest Neighbor (NN) and Manifold Projection (MP) methods outperform the unsupervised noisy RTF estimate, with MP showing consistently better performance on real data.
• The MP method provides superior blocking ability compared to NN in real-world conditions, especially under high reverberation and low SNR, demonstrating its robustness for practical applications.
• The database enables detailed analysis of RTF and RIR variations across a dense 3D grid, revealing subtle spatial response changes due to source position shifts.