Speaker
Description
Reconstructing sound-fields from a limited number of room impulse responses is a key challenge in spatial audio. Existing datasets are often spatially sparse or limited to a few rooms, hindering systematic benchmarking and training of sound-field reconstruction methods. This paper introduces Spatially Dense Room Impulse Responses (SpaDenRIR), a simulated dataset designed for developing and benchmarking sound-field reconstruction methods for early reflections. The dataset comprises 100 generated rooms, each containing a cubic microphone array of 22×22×22 microphones with 21.4 mm spacing, allowing spatial interpolation without aliasing up to 8 kHz. In every room, five sources are placed in distinct regions relative to the array. Room impulse responses are simulated using the image-source method and stored only for the first 100 ms, capturing the direct sound and early reflections. We describe the stochastic generation of room geometries, materials, source configurations, and array placement, and outline several application scenarios enabled by SpaDenRIR, including moving microphones and arbitrary array geometries. The dataset is intended as a shared training and evaluation resource for both model-based and machine-learning approaches to sound-field reconstruction.