eBrIRD - ELOSPHERES binaural room impulse response database

Overview

The ELOSPHERES binaural room impulse response database (eBrIRD) is a resource for generating audio for binaural hearing-aid (HA) experiments. It allows testing the performance of new and existing audio processing algorithms under ideal as well as real-life auditory scenes in different environments: an anechoic chamber, a restaurant, a kitchen and (forthcoming) a car cabin. The database consists of a collection of binaural room impulse responses (BRIRs) measured with six microphones. Two microphones represent the listener's eardrums. Four microphones are located at the front and back of two behind-the-ear hearing aids placed over the listener's pinnae. The database allows simulations of head movement in the transverse plane.

Technical Background

Hardware

The generation of the test signals, the calculation of BRIRs and their conversion into the SOFA format was accomplished in Matlab (R2019b) and the SOFA API for Matlab and Octave (V1.0). Signals were played from and recorded with a digital audio workstation (Cubase 10.5.11) connected to a soundcard (RME Fireface UFX II), while using the manufacturer's software (ASIO interface on MADIface version 0.9716; TotalMix FX version 1.63) running on a 'Windows 10 Professional' computer system. The soundcard allowed additional playback and recording over an ADAT slave (RME UCX, replaced by an RME Babyface in the anechoic scene). Active loudspeakers (Meyer MM-4XP) emitted the test signals that were captured by two microphones mounted inside the head of a manikin (GRAS KEMAR 45BB-1). A 2-channel power module (GRAS A12AR) polarised the capsules (GRAS RA0045) representing the eardrums and routed the signals to the microphone inputs of the ADAT slave. The signals from the front and back microphones inside two behind-the-ear HAs (Oticon Vigo) were sent to the four microphone inputs of the soundcard. A custom-build unit polarized these microphones. The HAs' receivers were taped on manikin's cheekbones, thus not blocking the ear canals in any way. A reference microphone (GRAS AF40) mounted on a pre-amplifier (Brüel and Kjaer 2260) and polarized with a one-channel power module (GRAS 12AK) was connected to the last microphone input of the soundcard during recording from the center-head position. The power module provided 10-dB gain and high-pass filtering (3-pole Butterworth, -1 dB @ 20 Hz). A sound-level calibrator (Brüel and Kjaer 4230) generated a 1-kHz reference signal at 94 dB SPL for all GRAS capsules. Their corresponding microphone pre-amplifiers were adjusted such that the reference signal was recorded with 3-dB headroom. The HA's microphones could not be calibrated, their gains on the microphone pre-amplifiers of the soundcard were set at equal levels, i.e., at 33 dB.

Audio

All audio was represented in 24-bits per sample at a rate of 44.1 kHz. Matlab-generated exponential sweeps spanning 62.5 to 16000 Hz had successive durations of 1, 2,4, and 8 s and 20-ms quarter-cosine shaped on- and offsets. A 3-s silence followed each sweep. A 5-s, 1-kHz pure tone tailed with 2 s of silence terminated the test signal. Sweeps and pure tones were played at 60- and 65-dB SPL, respectively, as measured under anechoic conditions with 1-m distance between the loudspeaker and the reference microphone. The test signal was subsequently played from each loudspeaker within an auditory scene and recorded with the six microphones. Each recording take provided the BRIRs for one head orientation, which was manually changed over a -90 to 90 degrees azimuthal range using 2.5 dB step. In a final take, the manikin was removed from the scene and the reference microphone was positioned at the place of the center of the manikin's head during previous takes. Each scene involved 74 takes. The original 8-s sweep was time-reversed, attenuated at a rate of 6-dB per octave and convolved with the original 8-s sweep as well as all recorded 8-s sweeps. This generated a band-limited Dirac delta function and large collection of BRIRs, respectively. The sample number of the peak in the Dirac delta function defined the start of the all BRIRs. Their samples prior to this sample number were deleted. We next determined the BRIR's noise floor per scene following Schroeder (1965) and set the BRIRs' length per scene to the lowest power of two such that the final BRIR included some but the smallest possible part of the noise floor.

Anechoic chamber

A first set of BRIRs were recorded in the anechoic chamber CGJ034 of the Laboratoire de Méchanique et d'Acoustique in Marseille, France. The manikin was placed in the middle of the chamber with two loudspeakers, i.e. one at 0° in front and another at 45° to the left. Both loudspeakers were at 1.07 m from the manikin's head center, at ear height and facing the manikin. The file 'Anechoic.sofa' captures the resulting BRIRs. The file AnechoicIR.sofa provides the impulse response of each loudspeaker as captured by the reference microphone.

Kitchen

Apartment A302, 50 traverse Parangon in Marseille, France provided the kitchen scene. The file Kitchen.sofa and KitchenIR.sofa contains the BRIRs and the room impulse response, respectively. Figure 1 illustrates the setup with the manikin seated at the kitchen table. LSP1 and LSP2 represented talkers seated on the other side of the table both facing the manikin. LSP3 placed on the counter imitates a kettle. The impulse response obtained from the reference microphone showed a reverberation time, the so-called RT60, of approximately 800 ms.

Figure 1: kitchen setup. Lay-out during the recordings in the kitchen scene. The room's height is 2.47m. The shaded area indicates the position of the kitchen table.

Restaurant

Room CGJ059a of the Laboratoire de Méchanique et d'Acoustique was used to simulate a small restaurant as illustrated in Figure 2. Twelve loudspeakers took positions of guests scattered around five tables. The manikin seated close to the center of the room shared a table with two loudspeakers: one in front and one to the left of our synthetic listener. The loudspeaker on the left pointed towards the loudspeaker in front. The latter loudspeaker pointed towards the manikin. Absorbent foam placed on the walls limited the reverberation time to around 470 ms, as calculated from the room impulse responses captured with the reference microphone. The file 'RestaurantIR.sofa' contains these impulse responses; the file 'Restaurant.sofa' the BRIRs.

Figure 2: restaurant setup. Lay-out during the recordings in the restaurant scene. The room's height is 3.17m. The shaded areas indicate table positions.

Car

In the car-cabin scene, the manikin took the front passenger seat of a Peugot 206+ constructed in 2009. The car was equipped with a Kenwood KMM-303BT digital media receiver connected to two Pioneer ts-1702i loudspeakers mounted in the doors. The impulse responses of the car's stereo system were measured while playing the test signals to the line input of the receiver with its volume set to 15 and any filtering deactivated. In addition, one active loudspeaker was placed behind the manikin on the rear seat and another at the driver's seat, simulating an additional passenger and the driver. All recording paths included 24 dB/oct high-pass filters provided by TotalMix with cut-off frequencies set at 62 Hz. The files 'Car.sofa' and 'CarIR.sofa' contain the BRIRs and IRs obtained with the reference microphone, respectively. Besides these impulse responses, we also recorded car-cabin noise. During these recordings the car's engine ran idle and its ventilation system was set at 2, all heating disabled and directed to head and feet. The right ventilation blower was closed, and all others pointed top left. We obtained 2-minute recordings for every head positions. These are stored in 'CarNoise.sofa', a file in SOFA format with the audio replacing the typical impulse responses.

Figure 3: car-cabin setup. Lay-out during the recordings in the car-cabin scene. The cabin's maxiumum height is 1.16m. Black elipses indicate the backs of the seats.

Download

Room impulse responses are stored in SOFA format.

Enquiries

Please mail enquiries to g.hilkhuysen@ucl.ac.uk.

Reference

Schroeder, M.R. (1965) New Method of Measuring Reverberation Time. J. Acoust. Soc. Am. 37, 409-12.


Some other pages on our site you may enjoy:

ESYNTH - Harmonic analysis/synthesis teaching tool

ESynth is a free program designed to explain the harmonic analysis and synthesis of signals. With ESynth you can create signals by adding together individual sinusoidal waveforms (sinewaves) and study the resulting waveform and spectrum. You can also perform an analysis of an input waveform, to see how a given sound can be represented in terms of a sum of sinewaves. More information.

RTSPECT - Real-time Waveform and Spectrum Display

RTSPECT is a free program for displaying a real-time waveform and spectrum display of an audio signal on Windows computers. More information.