[Paper Review] dEchorate: a Calibrated Room Impulse Response Database for Echo-aware Signal Processing

Diego Di Carlo, Pinchas Tandeitnik+4 more|arXiv (Cornell University)|Apr 27, 2021

Speech and Audio Processing61 references4 citations

TL;DR

dEchorate is a calibrated multichannel Room Impulse Response (RIR) database with precise annotations of early echo timings and 3D positions of sources, microphones, and image sources in a cuboid room under varying wall configurations. It enables benchmarking of echo-aware signal processing methods, achieving a mean distance error of 1.15 cm and angular error of 2.6° in room geometry estimation using 4 sources, demonstrating high accuracy for echo-based audio processing tasks.

ABSTRACT

This paper presents dEchorate: a new database of measured multichannel Room Impulse Responses (RIRs) including annotations of early echo timings and 3D positions of microphones, real sources and image sources under different wall configurations in a cuboid room. These data provide a tool for benchmarking recent methods in echo-aware speech enhancement, room geometry estimation, RIR estimation, acoustic echo retrieval, microphone calibration, echo labeling and reflectors estimation. The database is accompanied with software utilities to easily access, manipulate and visualize the data as well as baseline methods for echo-related tasks.

Motivation & Objective

• To provide a publicly available, highly calibrated RIR database with precise echo timing and spatial annotations for echo-aware audio signal processing.
• To support benchmarking of methods in acoustic echo retrieval (AER), room geometry estimation (RooGE), and speech enhancement (SE) under realistic conditions.
• To enable validation of robustness to variations in RT60, SNR, surface reflectivity, and early echo density.
• To facilitate research in microphone calibration, reflector estimation, and echo labeling using real-world measurements.

Proposed method

• The database was recorded in a controlled cuboid room with 11 microphone positions and 6 source positions, using calibrated transducers and precise 3D positioning.
• Early echoes were annotated using the Image Source Method (ISM), linking echo times of arrival (TOA) to virtual source positions relative to real reflectors.
• The dataset includes 1,800 RIRs across different wall absorption and reflection configurations, with accurate metadata on source, microphone, and wall positions.
• A software toolkit was developed to access, visualize, and manipulate the data, including baseline implementations for AER and RooGE tasks.
• The RIRs were measured using exponentially swept-sine (ESS) signals to ensure high signal-to-noise ratio and accurate impulse response estimation.
• Outliers in echo annotations were analyzed and attributed to directivity effects, far-field assumption violations, and overlapping reflections.

Experimental results

Research questions

• RQ1 How accurately can echo-aware signal processing methods estimate room geometry using real-world RIRs with annotated early echoes?
• RQ2 To what extent do mismatches between simulated and real RIRs affect the performance of learning-based audio processing models?
• RQ3 Can robust echo labeling be achieved in the presence of low-energy reflections, overlapping echoes, or directivity effects?
• RQ4 How do variations in surface absorption, RT60, and microphone-source proximity impact the reliability of echo-based processing?
• RQ5 Can the simulated-to-real RIR mapping be learned via domain adaptation or style transfer techniques?

Key findings

• The room geometry estimation pipeline achieved a mean distance error (DE) of 1.15 cm and angular error (AE) of 2.6° using data from 4 sources, demonstrating high accuracy.
• Out of 1,800 RIRs, only two echo labeling outliers were observed, both due to source directivity causing misannotation of first-order reflections.
• The dataset revealed that first-order reflections behind sources can be too weak to appear in RIRs, leading to second-order images being incorrectly selected.
• Late reverberation spikes caused by long impulse responses in microphones and loudspeakers led to misclassification of echo components in some cases.
• The dataset supports robust benchmarking across varying RT60, SNR, and echo density, enabling evaluation of method robustness.
• The availability of paired simulated and real RIRs opens the door to developing domain adaptation techniques for learning-based acoustic simulators.