gms | German Medical Science

In everyday life, complex hearing situations with background noise often occur. In addition, in closed rooms direct sound is superimposed with reverberation. Typical listening situations in everyday life are, for example, babble noise of many speakers in a restaurant, at the train station or in a supermarket. Speech perception is highly dependent on the number and spatial arrangement of noise sources [1]. Test procedures in the diagnosis and therapy of hearing disorders as well as in the fitting and testing of hearing systems are usually performed in acoustically optimized listening booths with very low reverberation times. Therefore, the actual hearing performance in everyday life cannot be reproduced. Here, only one or two sound sources (loudspeakers) are used, which limits the possible directions of incidence of target sound (i.e. speech) and noise. Simulation methods offer the possibility to simulate everyday listening situations in order to examine the hearing performance of people with and without hearing loss [2], [3], [4].

The sum of all reflections within a delimited room is defined as reverberation [4]. The acoustic information produced by a sound source in closed rooms is superimposed by these reflections. The signal that arrives at the receiver can be divided into direct sound, early and late reflections. Early reflections occur up to 50–80 ms after the first sound wave and are partly beneficial for speech perception because they provide amplification of the information of the direct sound [6]. Late reflections occur with more delay and are perceived diffusely or as a separate echo with the consequence of deteriorated speech perception. The acoustic properties of a room are expressed by various measures such as reverberation time (RT), clarity measures (C50/C80), “Deutlichkeit”/definition (D50) or strength (G). The reverberation time describes the time it takes for the sound pressure level to decrease to one thousandth of its initial value (i.e. by 60 dB). The clarity C50 describes the ratio between the sound energy in the period up to 50 ms after direct sound and the sound energy from 50 ms onwards. Clarity C80 is defined by the ratio between the sound energy up to 80 ms after the direct sound and the sound energy from 80 ms onwards. The definition D50 describes the energy ratio between the early reflections of the first 50 ms to the total sound energy. The strength G puts the sound level of a spherically radiating sound source in relation to the level of the same source at a distance of 10 m. The mentioned room acoustic parameters are influenced by the size (i.e. volume) of the room and the acoustic properties of the boundary surfaces such as ceilings, floor, walls, etc. Depending on the acoustic properties of a room, the temporal and spectral structure of the reproduced signal is changed. The superposition of direct sound with reverberation can be compared to the effect of a masking noise and can additionally reduce the modulation depth of the target signal.

Due to the dependency between room size/condition and speech perception, DIN 18041:201603 (audibility in rooms) recommends suitable reverberation times depending on the volume of the room [7]. Depending on the room usage, there are different recommendations for music, speech, teaching, etc. For room volumes between 1,000 to 5,000 m³, the optimum reverberation time for teaching is approx. 0.75 to 1 s, rooms for teaching lessons for a handicapped audience should not exceed 0.6 to 0.8 s reverberation time for the specified room volume.

This study describes the further development of an already existing free field reproduction system with 128 loudspeakers [8] for reproducible, flexible and plausible reproduction of sounds in rooms with different acoustic properties. The evaluation includes the analysis of room acoustic parameters and interaural level differences in three reverberated conditions. Furthermore, the influence of reverberation on speech perception in noise will be investigated under consideration of spatial release from masking (SRM) in normal hearing subjects.

For the implementation of the simulation system, an anechoic chamber was available at the Audiological Acoustics Department of the ENT University Hospital of the J.W.-Goethe University Frankfurt am Main. This room is equipped with a reproduction system with 128 loudspeakers (rectangular array in the horizontal plane), which was previously used either under free field conditions (e.g. [8], [9], [10]) or for the direction-independent simulation of diffuse sound [11]. The existing system was extended by a component for room simulation.

The software package ODEON (ODEON A/S, Lyngby, Denmark) was used with to implement a three-dimensional model of an empty lecture hall (see Figure 1 [Fig. 1]) with a room volume of 3,520 m³. The height of the room is between 7.5 to 10 m. This room model is freely available and delivered as example with ODEON (file name: Example.par). In ODEON, absorption properties or specific material properties can be assigned to objects and walls. Based on the room geometry mentioned above, two room models with different absorption properties of all room surfaces (i.e. walls, ceiling, floors) were generated. The frequency independent absorption coefficients were 80% and 60%. Thereby, 80% absorption corresponds to an average reverberation time of 350 ms and 60% to a reverberation time of 510 ms.

Three sound sources were positioned in the front half space of the horizontal plane at –60°, 0° and 60° relative to the receiver. The frontal sound source (0°) was 5 m apart from the listening position, the lateral sound sources were 10 m apart from the listening position. Figure 1 [Fig. 1] (left) shows a sketch of the room geometry of the room with three sound sources and one receiver. In the room model, the sound propagation of a signal emitted by the transmitter independently of direction and frequency was calculated (mirror source model and “ray-radiosity method”, [12]). Reflections up to the 10^th order were calculated and stored in a reflectogram. For each reflection, the reflectogram contains the time delay of the reflections, the octave levels between 63 and 8,000 Hz, the horizontal sound incidence angles and the vertical sound incidence angles. Figure 1 [Fig. 1] (right) displays sound propagation. During its propagation, the radiated sound hits room boundaries several times and is reflected repeatedly. The direct sound is shown in dark red and first-order reflections in green.

Based on the reflectograms calculated by ODEON, the room simulation was realized in a low-reflective room. Figure 2 [Fig. 2] shows a sketch of the laboratory room with the positions of the sound sources.

The nearest loudspeaker method was used [13], [14]. This means that for each calculated reflection, the speaker with the smallest distance of the horizontal angle is selected for playback. Reflections with an elevation angle of incidence greater than ±30° were not taken into account because the speakers of the present sound reproduction system are arranged horizontally. While a complete integration (i.e. mapping) of all elevation angles in the horizontal plane allows a complete preservation of the reflection density of the initial reflections, this variant leads to clearly audible sound coloration. After preliminary listening experiments, vertical angles of incidence up to ±30° were discarded as a compromise between reflection density and sound coloration. Based on the reflectogram, an impulse response was generated for each reflection according to its time delay and spectrum.

In the next step, each reflection was assigned to the loudspeaker closest to the calculated horizontal angle and the impulse responses per loudspeaker were summed. Late reflections were inserted as diffuse reverberation using a statistical model 80 ms after direct sound incidence (Feedback Delay Network, FDN retrieved from Quality & Usability Labs, Institution for Software Technology and Theoretic Informatics, Technical University Berlin). For this purpose, the RMS levels of all generated reflections were determined in the time window between 80 ms and 120 ms after direct sound. The level of the diffuse reverberation was adjusted in level to the previously generated reflections. The diffuse reverberation was generated using the frequency-dependent RT and faded in equally on all 128 channels. To compensate for the individual transmission characteristics of the loudspeakers to the center of the room (amplitude frequency response and running time) due to the rectangular arrangement, equalization was carried out individually for each loudspeaker (finite impulse response (FIR) filter). The generated 128-channel room impulse response (RIR) was convolved with the desired single-channel audio signal for sound reproduction. The spatial sound was presented simultaneously via all 128 loudspeakers. RIRs of both conditions with reverberation were generated, each for three direct sound positions (±60°, 0°). Figure 3 [Fig. 3] shows a flowchart of the room simulation with direct sound, early reflections and diffuse reverberation.

To check the quality of the simulation, RT and D50 were measured and compared with the model data in the first step. In a second step, the system was evaluated from an audiological point of view by conducting speech tests and evaluating the results. The influence of the reverberation time on speech perception and SRM was investigated. In addition, head-related IRs (HRIRs) recorded with an artificial head were examined for interaural level differences (ILDs) and considered in relation to the SRM.

The RT and D50 were analyzed per octave frequency band from 63 Hz to 8,000 Hz as well as broadband averaged and compared with the calculations of the room model. The analysis was performed for a sound source at 0°. A sinus sweep (frequency range from 50 Hz to 22,050 Hz) of 0.5 s duration was used as stimulus. The measurements at the listening position were made with a measuring microphone of type 4155 (Brüel & Kjær, Nærum, Denmark) and an impedance converter 2,669 connected to a measuring amplifier (Nexus) of the same manufacturer. The impulse responses were analyzed with the acoustic software ARTA (Artalabs, Kastel Luksic, Croatia). In addition, the IRs were measured at the head and torso simulator (Brüel & Kjær Type 4100 with microphones Type 4190-L-002). The sound presentation was made from 60°. ILDs were derived from the results of the measurements. All measurements were performed in both simulated conditions with absorption coefficients of 60% and 80% and under free field conditions (i.e. ~100% absorption).

For the free field measurement an average RT of 0.05 s was obtained. The RT values per octave band are given in Table 1 [Tab. 1] for measurement condition 2 with 80% absorption and measurement condition 3 with 60% absorption. The RT averaged over all frequencies determined by the model using ODEON was 0.35 s in the condition with 80% absorption and 0.51 s with 60% absorption. The measurements with the simulation system in the (nearly) anechoic room resulted in a reverberation time of 0.39 s at 80% absorption and 0.58 s at 60% absorption. The greatest deviations were found in the lower frequency bands, especially in the range and below the cut-off frequency of the low-reflective room (~200 Hz).

Table 2 [Tab. 2] shows the D50 measurement per octave frequency as an example for measurement condition 2 (80% absorption) and measurement condition 3 (60% absorption). In the model, the averaged D50 values were 95% and 91% over all frequencies (condition 2 and 3 accordingly). The measurements showed 96% in condition 2 and 87% in condition 3. Similar to the RT values, the largest deviations are in the lower frequency range. Using the HRIRs recorded on the head and torso simulator, ILDs were calculated for all three conditions. Figure 4 [Fig. 4] shows the ILDs as a function of frequency. In all conditions the highest ILD is at about 4 kHz. Up to 4 kHz the ILDs rise steeply with increasing frequency, from about 5 kHz they decrease slightly and remain almost constant up to 10 kHz. The stronger the reverberation, the lower the ILD. Between condition 1 (free field condition) and condition 3 (strong reverberation) there is an ILD difference of up to 10 dB between 3,000 Hz and 5,000 Hz.

Speech perception in noise was determined in normal hearing (NH) with the Oldenburg sentence test (OLSA, [15], [16], [17] in the source arrangements S0N0 (signal at 0°, noise at 0°) and S0N60 (signal at 0°, noise at ±60°).

The distance between the subject and the loudspeaker at 0° was 1 m, to the loudspeakers at 60° the distance was 1.80 m. The direct sound of the speech signal was set at 60 dB SPL for all conditions. The level of the speech signal remained constant, while the noise level was adaptively changed. The initial signal to noise ratio (SNR) was set to +5 dB. All subjects completed a training list. Subsequently, measurements with 100%, 80% and 60% absorption were performed in randomized order with the S0N0 and S0N60 setups. As noise signal the Oldenburg noise (OlNoise) was used. One run consisted of 20 sentences. The subject was placed in the center of the test room and responded via a touch screen monitor.

17 NH subjects (age: 26.7±8.0 years, 3 female, 14 male) participated in the measurements. The results were examined for normal distribution using the Kolmogorov-Smirnov test. Since the test confirmed the assumption of a normal distribution, statistical evaluations were carried out using the T-test. In multiple comparisons, the significance level was corrected using the Bonferroni method.

Figure 5 [Fig. 5] shows boxplots of the speech perception thresholds (SRTs) for the three conditions in the two source arrangements S0N0 (Signal 0°, Noise 0°) and S0N60 (Signal 0°, Noise ±60°).

For speech perception with spatially superimposed speech and noise playback from the front (S0N0), the free field condition resulted in a median SRT of –7.4 dB SNR. In condition 2 this is reduced to –5.8 dB SNR. The results are significantly different (p=0.024). In condition 3 the SRT is still –4.7 dB SNR. This is significantly worse than the FF condition (p<0.001) and than in the condition with 80% absorption (p=0.01).

With spatial separation of the speech and noise sources (S0N60), the SRT in the free field condition improved to –13.9 dB SNR. At an absorption coefficient of 80%, the SRT is reduced to –8.8 dB SNR, at an absorption coefficient of 60% it is still –6 dB SNR. All results differ significantly (p<0.001).

The benefit in terms of SRT due to spatial separation (SRM) was determined by subtracting the SRT of conditions S0N60 and S0N0. SRM results are shown in Figure 6 [Fig. 6]. In the free field, the advantage of spatial unmasking is 6.8 dB. With 80% absorption, the SRM is reduced to 2.7 dB; with 60% absorption it is only 1.3 dB. The results of conditions with reverberation differ significantly from free field condition (p<0.001) and the SRM at 60% absorption is significantly lower than at 80% absorption (p=0.016).

The aim of this work was the extension of our custom free field reproduction system by a room simulation procedure and the subsequent evaluation of the setup. For this purpose, room simulations were performed for a model auditorium and the acoustic parameters RT and D50 were compared between the predictions of the model and the laboratory reproduction. In addition, the effect of two different reverberation times on speech perception in noise and SRM was tested by introducing two different absorption levels.

Similar to other works, a room simulation based on reflectograms from a room model was realized [2], [13], [14] and supplemented by the diffuse reverberation using a feedback delay network. For the reproduction of the initial reflections, the approach of the nearest loudspeaker was chosen. This method was used because the integration into the already existing test setup with 128 loudspeakers could be done without technical modifications of our local reproduction room. Furthermore, the small diameter of the individual loudspeakers (and the resulting small loudspeaker distance) allow an extremely good angular resolution in the horizontal plane. Other systems with fewer loudspeakers, which have already been described in the acoustic room simulation literature, apply “Higher Order Ambisonics” (HOA, e.g. [2]), Vector Base Amplitude Panning (VBAP, [18]) or Wave Field Synthesis (WFS, [19]). Methods such as HOA and WFS have the disadvantage, that due to the limited number of loudspeakers or due to the not infinitely small distance between two loudspeakers, a more or less pronounced spatial subsampling (so-called “spatial aliasing”) occurs, i.e. the synthesized sound field is no longer physically correct above the alias frequency. With WFS, the alias frequency is usually around 1–4 kHz, depending on the arrangement and speaker distance. The method of the nearest loudspeaker used here was favored, since it always allows physically correct reproduction of the individual reflections in the entire relevant frequency range. The disadvantage of the system described here is that only a playback in the horizontal plane is possible. To enable a three-dimensional room simulation, the use of HOA or VBAP is recommended. To accomplish this, loudspeakers with elevated positions must be installed. If the listener position is held constant in the so-called “sweet spot” (center of the playback array), a sufficiently correct spatial resolution for hearing research can also be achieved when using HOA with the appropriate playback setup. In the three-dimensional implementation, the diffuse reverberation can also be generated with 1^st order ambisonics based on the envelope [2]. Instead of using a room model, three-dimensional microphone recordings (i.e. measured impulse responses) of real rooms can also be used for room simulation with the HOA reproduction method.

While the method used here cannot generate dynamically moving sound sources, this is possible with other methods. Depending on the problem and the desired listening situation, the choice of the reproduction method used must therefore be considered individually with the respective advantages and disadvantages. The studies of Grimm et al. [20] and Ahrens et al. [21] provide a good overview of which methods are best suited for which problem.

The evaluation of the room simulation presented here showed a good agreement between measurements and model data. Both RT and D50 showed larger deviations only below the cut-off frequency of the local anechoic chamber. The auditorium simulated in this study had a volume of 3,520 m³ and according to the measurements a maximum RT of 580 ms.

The reverberation times accomplished in the present study were well below the recommended maximum average reverberation time of 900 ms for teaching/communication (1.2 s for speech/lecture) as required by the DIN standard [7]. Nevertheless, speech intelligibility was already severely impaired.

The results on speech perception in source arrangement S0N0 condition 3 (60% absorption, RT 580 ms) showed a deterioration of 2.7 dB SNR against FF situation. In condition 2 (80% absorption, RT 390 ms), the deterioration was 1.6 dB SNR. Due to early reflections, the emitted sound reaches the ear several times with small delay, additionally late reflections cause the sound level to decrease slowly. In speech listening tasks, this leads to a masking of speech components by sound reflections. Individual phonemes may be masked, which leads to reduced speech perception. The RT and D50 also express this: while in condition 1 (free field condition) there is a low RT and 100% of the sound energy is integrated within 50 ms, in condition 3 an average of 13% of the sound energy occurred after the first 50 ms after direct sound incidence. This suggests an increased masking of speech by reverberation.

In the S0N60 source setup at 60% absorption, the SRT deteriorated by 7.9 dB SNR compared to the FF condition. This was particularly reflected in the SRM. Only 1.3 dB SNR improvement could be achieved by spatial separation of noise and target sources. Figure 4 [Fig. 4] shows the ILDs of the three measured acoustic conditions as a function of frequency. Between condition 1 (free field) and condition 3 (60% absorption), reduced ILDs are shown over the entire frequency spectrum. Between 3,000 Hz and 5,000 Hz the ILD difference is at its maximum with about 10 dB. This ILD reduction explains the significantly reduced SRM under reverberation, the shadowing of the sound by the head is reduced in a diffuse sound field. The SNR improvement in the contralateral ear caused by the head shadow effect is therefore significantly lower and results in a reduced effect. Kidd and co-authors [18] demonstrated the effect of reverberation on the SRM in a room lined with plexiglass in one condition and foam in the other. The benefit of spatial separation of target sound and noise by 90° decreased from 8 dB SNR to 2 dB SNR in the plexiglass condition and is in line with the results shown here.

The results confirm the long known observation of the impact of room acoustics on speech perception. Room acoustics with very small reverberation time promote good speech perception. A long reverberation time, on the other hand, leads to increased masking of the direct speech signal by additional reverberation components of the background noise and the target signal itself. Long reverberation times therefore have a negative effect on speech perception. The influence of room acoustics is currently not taken into account in clinical audiological test procedures for speech perception. The determination of hearing performance by means of a room simulation system is a useful supplement to established audiological measurement procedures. In the future, hearing tests with hearing aid and cochlear implant users should be performed in simulated acoustic listening environments. This will provide additional information on the impact of the type of hearing aid and the effect of signal pre-processing on speech perception. A supplement to the measurement conditions is planned and should lead to further insights into the effect of room acoustic parameters on speech perception. Especially the investigation of classrooms with different room acoustic properties and different distances between sound source (teacher) and listening position (student) are of interest. The influence of room acoustics on directional hearing will also be investigated. Thanks to room simulation, listening situations can now be reconstructed in the laboratory under realistic but controlled acoustic conditions.

The authors declare that they have no competing interests.

The authors thank the company Cochlear Deutschland GmbH & Co. KG and the Moessner Foundation Frankfurt am Main for supporting the study.