The objective of this study is to determine the minimum level of noise in a restaurant that starts the Lombard effect, and how it relates to the perceived communication disturbance and the willingness to spend time and money for a meal. Twenty-eight participants were instructed to read a passage in the presence of restaurant noise from 35 to 85 dB(A). As the noise level increased, participants began to be disturbed by the noise at 52 dB(A) and began to raise their voice at 57 dB(A). The willingness to spend time and money decreased starting at 52 dB(A).
1. Introduction
According to the 2016 Zagat State of American Dining report,1 25% of restaurant customers consider noise the most irritating component of dining out. Restaurants are not merely eating sites, they also serve as business and social hubs. Typical noise levels in restaurants range from 60 to 80 dBA.2 However, Rohrmann3 found that many restaurants exceed these levels.
In a noisy environment, speakers unconsciously attempt to maintain a level of speech that allows them to be understood (Lombard effect). Many studies have addressed the Lombard effect in a variety of conditions.4 Various slopes of the relationship between the noise level and voice level have been reported. These slopes vary according to boundary conditions such as the speech situation (reading vs conversing), type of noise (machinery noise, office noise, restaurant noise, speech noise, wideband noise, white noise, pink noise), style of speech (normal or shouting), speaker-listener distance, and room acoustics.5 Lazarus6 stated that the speech level rises with noise level with a slope of 0.3–0.6 dB per noise level rise of 1 dB for all disturbing noise exceeding 40–50 dB(A). However, up to a noise level of 30–40 dB(A), the noise level has minimal effect on the speech level.7
Prior studies have used different types of background noise (classical music vs restaurant noise) within a variety of eating establishments (cafe vs restaurant) to examine the correlation between comfort levels of the patrons and the satisfaction of their experience. In a restaurant setting, the Lombard effect is due more to the intensity of conversation held by patrons rather than other background noise.8 Other research reports suggested acoustical capacity in the restaurant setting, in regards to how many patrons can comfortably maintain verbal communication in an establishment.9 Any intensity level which surpasses this “comfortable” range can lead to a loss of customer satisfaction and a decrease in business for the establishment.10,11
Tang et al.12 estimated the Lombard effect by measuring the variation of noise level with the number of occupants in a university staff canteen. A method, which takes into account the effect of raising voices, was derived to predict the variation of A-weighted sound level inside an enclosure. From their results, they deduced that occupants began to raise their voices when the background noise level exceeded 69 dB(A). Hodgson et al.13 proposed a model for predicting speech and noise levels, and the acoustical conditions for verbal communication including the Lombard effect, in eating establishments. The Lombard slope resulting from the proposed model was 0.69 dB/dB.
In sum, although close attention has been paid to the Lombard effect, information about the existence of the starting point, in terms of noise level, and the slope in restaurant settings is still based on estimations.
The aim of this study is to identify the specific point when the restaurant noise causes vocal discomfort for customers. The study also seeks to discover how willing a customer would be to spend their time and money in a restaurant, depending on the varying intensities of noise levels in the surrounding environment.
The main research questions of this study were:
Is there a starting-point in the level of the noise for the Lombard effect in restaurant settings?
Is there a starting-point for self-reported communication disturbance in restaurant settings, and how does this relate to the starting-point for the Lombard effect?
Is there an effect of the background noise level on the time and the budget we would spend in a restaurant?
It was hypothesized that there would be a starting point for the Lombard effect ranging from 40 to 50 dB(A). This range was considered appropriate due to the results procured from previous studies.7 Overall, the relationship between the starting point of the Lombard effect in a restaurant setting and the customer's vocal discomfort, along with their willingness to spend money and time at the establishment was investigated.
2. Experimental method
In this study, the effects of typical restaurant noise varying between 35 and 85 dB(A) on (1) vocal effort [quantified as sound pressure level (SPL)], (2) the amount of self-reported disturbance in the communication due to noise and vocal control, and (3) the willingness to spend time and money for a meal, were evaluated.
With protocol approval of the University of Illinois Urbana-Champaign's Office for the Protection of Research Subjects Review Board (IRB No. 18179), 14 male and 14 females were recruited to participate. The speech of the 28 talkers (18–28 yr; mean 21 yr) was recorded in a sound-attenuated booth with typical restaurant noise. These participants reported no history of a speech impairment and they were audiometrically assessed between 250 Hz and 6 kHz to confirm normal hearing lower than or equal to 20 dB hearing level.
In order to simulate a real communication setting, participants were seated in the booth facing a human listener, positioned at a 1 m distance. The 10 noise conditions (see Sec. 2.1) were presented in a random order. Participants were asked to read the text to the listener using the following instructions: “Each time, I [the listener] would like you to pretend that we are talking in a restaurant and you are telling the story to me. Make sure that I understand you equally well each time.” The listener was present in the booth during the entirety of the experiment.
The participants were instructed to read a 6-sentence excerpt from the Rainbow passage.14 After each reading of the text, participants were asked to answer three questions about the experience of talking in the various noise level conditions. The questions were as follows:
Disturbance: Please rate the amount of disturbance you perceived during your communication by noise. (The extremes of the lines were “very low” to the left and “very high” to the right.)
Time: How long would you enjoy your stay at this restaurant? (The extremes of the lines were “No time” to the left and “A long time” to the right.)
Budget: How much of your budget would you spend in this restaurant? (The extremes of the lines were “None of it” to the left and “All of it” to the right.)
Participants responded to the questions by making a vertical tick on a continuous horizontal line of 100 mm length (a visual analog scale) immediately after the exposure to the noise of the task.
2.1 Room acoustic and measurement procedures
The experiment took place in a sound-attenuated booth (2.1 m × 2.2 m × 2.0 m). The speech was acquired by a head-mounted microphone (Beta 54 WBH54, Shure, Niles, IL) and connected to a PC via an audio interface (US-20x20, Tascam, Montebello, CA).
The speech was recorded with the software Audacity 2.0.6 (SourceForge, La Jolla, CA) in 11 noise conditions between 35 and 85 dB(A) in 5 dB increments. The noise levels for the 11 conditions were measured with a measurements microphone M2211 (Class 1 frequency response) and analyzed by means of an Audio and Acoustic Analyzer [level range 10–110 dB(A)], produced by NTi Audio (Schaan, Liechtenstein). The measurements were performed by placing the microphone in the position of the participants' ears. Typical restaurant noise was emitted by two directional speakers (studio monitor model Rokit5 G3, produced by KRK System, Chicago, IL) placed at 45°, 1.5 m in front of the participant. The gain of the playback software of the studio monitors was modified in order to obtain increments of 5 dB in the position of the participants' ears. The spectrum of the restaurant noise during the experiment is shown in Fig. 1.
Relative spectrum of the restaurant noise used during the experiment, per one-third octave band.
Relative spectrum of the restaurant noise used during the experiment, per one-third octave band.
Reverberation time was measured in the sound booth from the impulse responses (IRs) generated by balloon pops.15 The four IRs were recorded in two source positions and two microphone positions by means of an NTI Measurements microphone M2211 (Class 1 frequency response) and analyzed in one-third octave bands by means of an NTI XL2 Audio and Acoustic Analyzer. The reverberation time (T20) at mid-frequencies in the room was 0.05 s, while the background noise was 22.5 dB(A).
2.2 Analysis
Matlab (R2017a) was used for speech signal analysis. In each condition, the equivalent SPL was measured
For each condition, the mean value of the SPL was obtained per subject. For each subject, the average of SPL among the conditions was computed and subtracted from each mean SPL values for that subject (termed ΔSPL). This within-subject centering was performed in order to evaluate the variation in the subject's vocal behavior in the different noise conditions from their typical vocal behavior (mean value of the SPL per subject).
The levels were a combination of two sources: the voice and the noise. In order to evaluate the Voice to Noise Ratio (VNR) in the recordings, the distributions of the two sources were studied using the Expectation-Maximization algorithms for Gaussian mixtures. The algorithm allows analyzing the mixture of distributions. In our case, the distribution of sound levels is a mixture of the voice and the noise levels. The algorithm estimates the mean values of the two distributions. The difference between the two mean levels represents an estimation of the VNR. The analysis was performed on a time history of the SPL, with a time step of 0.05 s, considering the subset of the dataset per noise condition.
Self-reported communication disturbance and willingness to spend time and money were measured on visual analog scales. The score was measured as the converted percentage of the distance of the tick from the left end of the line.
Statistical analysis was conducted using R version 3.1.2. Four piecewise linear (also called segmented or broken-line) models were fit to the response variables SPL, self-reported discomfort, and willingness to spend time and money with the predictor, noise level, using the segmented package in R. In such models, the fitted lines are constrained to be connected at the estimated change-point, i.e., the change-point in the relationship between the response variable and the predictor variable. At the change-point, it is assumed that the mean of the parameter is constant between the two slopes. If the first slope is equal to zero, the change-point can be considered a starting point. First, a simple linear model is fit. Subsequently, using the segmented function, maximum-likelihood methods are used to determine the slopes of the regression lines and the location of the change-point. No initial guess for change-point locations or the number of change-points is supplied. The confidence intervals for the change-point are estimated using the standard error from the Delta method for the ratio of two random variables.16 Because the segmented function16 accepts input only as simple linear models and not mixed-effect models, the between-subject variability was taken into account by the within-subject centering.
3. Results
As a first result, the VNR in the recordings was evaluated. The average VNR among the different noise conditions was 22.1 dB with a standard deviation of 1.3 dB. This result confirms that the effect of noise on the equivalent level was negligible. In the worst case (VNR = 20.0 dB), the contribution of the background noise on the overall level (noise and voice) was about 0.1 dB.
The piecewise linear (also called segmented or broken-line) models appear to be the best fits in comparison to simple linear models and quadratic models. The goodness of the fit was evaluated based on the R-squared, the analysis of the residuals, and the fact that the use of a more complex model did not improve the fit in a statistically significant way.
The ΔSPL was measured at each of the 11 noise levels between 35 and 85 dB(A), as shown in Fig. 2. A piecewise linear model was fit to the response variable, ΔSPL, and the predictor, Ln. The slope of the lower segment was 0.31, and the upper was 0.54, with a change-point identified in Ln at 57.3 dB(A) (confidence interval, CI 95% lower: 53.0, CI % upper: 61.6) with an R-squared of 0.91. Model estimates with associated standard errors and p-values are given in Table 1.
Relationship between the level of the noise in dB(A) and self-reported communication disturbance (a), relative voice level (b), willingness to spend time (c), and willingness to spend money (d), where the error bands indicate the standard error. Vertical dashed lines mark the change-points.
Relationship between the level of the noise in dB(A) and self-reported communication disturbance (a), relative voice level (b), willingness to spend time (c), and willingness to spend money (d), where the error bands indicate the standard error. Vertical dashed lines mark the change-points.
Piecewise linear model output for four models with response variables ΔSPL, disturbance, time, and budget as a function of Ln.
| Response | Predictor | Estimate | Std. Error | t value | p | Ln Domain / dB(A) |
|---|---|---|---|---|---|---|
| ΔSPL / dB(A) | (Int.) | −20.76 | 1.11 | −18.68 | <0.001 | 35 ≤ Ln ≤ 57.3 |
| Ln | 0.31 | 0.02 | 12.90 | <0.001 | ||
| (Int.) | −33.77 | 1.46 | −23.05 | <0.001 | 57.3 < Ln ≤ 85 | |
| Ln | 0.54 | 0.02 | 26.99 | <0.001 | ||
| Disturbance/% | (Int.) | −22.75 | 8.09 | −2.81 | 0.006 | 35 ≤ Ln ≤ 52.2 |
| Ln | 0.74 | 0.19 | 3.91 | <0.001 | ||
| (Int.) | −115.51 | 8.97 | −12.87 | <0.001 | 52.2 < Ln ≤ 85 | |
| Ln | 2.51 | 0.13 | 19.81 | <0.001 | ||
| Time/% | (Int.) | 112.98 | 9.42 | 11.98 | <0.001 | 35 ≤ Ln ≤ 51.3 |
| Ln | −0.65 | 0.22 | −2.98 | 0.003 | ||
| (Int.) | 194.53 | 8.30 | 23.43 | <0.001 | 51.3 < Ln ≤ 85 | |
| Ln | −2.24 | 0.12 | −19.12 | <0.001 | ||
| Budget/% | (Int.) | 99.93 | 13.05 | 7.65 | <0.001 | 35 ≤ Ln ≤ 52.5 |
| Ln | −0.63 | 0.30 | −2.06 | 0.04 | ||
| (Int.) | 157.44 | 8.99 | 17.51 | <0.001 | 52.5 < Ln ≤ 85 | |
| Ln | −1.72 | 0.13 | −13.55 | <0.001 |
| Response | Predictor | Estimate | Std. Error | t value | p | Ln Domain / dB(A) |
|---|---|---|---|---|---|---|
| ΔSPL / dB(A) | (Int.) | −20.76 | 1.11 | −18.68 | <0.001 | 35 ≤ Ln ≤ 57.3 |
| Ln | 0.31 | 0.02 | 12.90 | <0.001 | ||
| (Int.) | −33.77 | 1.46 | −23.05 | <0.001 | 57.3 < Ln ≤ 85 | |
| Ln | 0.54 | 0.02 | 26.99 | <0.001 | ||
| Disturbance/% | (Int.) | −22.75 | 8.09 | −2.81 | 0.006 | 35 ≤ Ln ≤ 52.2 |
| Ln | 0.74 | 0.19 | 3.91 | <0.001 | ||
| (Int.) | −115.51 | 8.97 | −12.87 | <0.001 | 52.2 < Ln ≤ 85 | |
| Ln | 2.51 | 0.13 | 19.81 | <0.001 | ||
| Time/% | (Int.) | 112.98 | 9.42 | 11.98 | <0.001 | 35 ≤ Ln ≤ 51.3 |
| Ln | −0.65 | 0.22 | −2.98 | 0.003 | ||
| (Int.) | 194.53 | 8.30 | 23.43 | <0.001 | 51.3 < Ln ≤ 85 | |
| Ln | −2.24 | 0.12 | −19.12 | <0.001 | ||
| Budget/% | (Int.) | 99.93 | 13.05 | 7.65 | <0.001 | 35 ≤ Ln ≤ 52.5 |
| Ln | −0.63 | 0.30 | −2.06 | 0.04 | ||
| (Int.) | 157.44 | 8.99 | 17.51 | <0.001 | 52.5 < Ln ≤ 85 | |
| Ln | −1.72 | 0.13 | −13.55 | <0.001 |
Self-reported communication disturbance was measured at each of the 11 noise levels (Fig. 2). A piecewise linear model was fit to the response variable, disturbance (% of very high) and the predictor, Ln. The slope of the lower segment was 0.74%-points per dB (pp/dB), and the upper was 2.51 (pp/dB), with a change-point in Ln identified at 52.2 dB(A) (CI 95% lower: 48.0, CI % upper: 56.4) and an R-squared of 0.80. Model estimates with associated standard errors and p-values are given in Table 1.
The willingness to spend time in the restaurant was measured at each of the 11 noise levels, as shown in Fig. 2. A piecewise linear model was fit to the response variable, time willing to stay (% of A long time) and the predictor, Ln. The slope of the lower segment was −0.65 (pp/dB), and the upper was −2.24 (pp/dB), with a change-point in Ln, identified at 51.3 dB(A) (CI 95% lower: 46.9, CI % upper: 55.7) with an R-squared of 0.77. Model estimates with associated standard errors and p-values are given in Table 1.
The willingness to spend money in the restaurant was measured at each of the 11 noise levels, as shown in Fig. 2. A piecewise linear model was fit to the response variable, money willing to spend (% of “All the budget”) and the predictor, Ln. The slope of the lower segment was −0.63 (pp/dB), and the upper was −1.72 (pp/dB), with a change-point in Ln, identified at 52.5 dB(A) (CI % lower: 44.7, CI % upper: 60.2) with an R-squared of 0.60. Model estimates with associated standard errors and p-values are given in Table 1.
4. Discussions and conclusions
Based on the reported results, which are comparable to previous work, it can be claimed that vocal level (effort) and disturbance increase as background noise increases, while the willingness to spend time and money in a restaurant decreases as background noise increases.
The hypothesis of a starting point for the Lombard effect was not verified. However, as background noise increased, change-points could be identified in the slope of the increase in vocal level (Lombard effect), disturbance, and willingness to spend time and money in a restaurant. Regarding the objective measure of vocal effort, a change-point of the Lombard effect was identified at a noise level equal to 57.3 dB(A), while Tang et al.12 deduced that occupants began to raise their voices when the background noise level exceeded 69 dB(A). Previous studies showed a starting point for the Lombard effect around 40–50 dB(A).6,7
The difference in the changing point level could be due to different factors: in the present study the participant and the research assistant were the only people in the room, this configuration probably inhibited possible automatic mechanisms in speech regulation related to privacy. This could be an explanation for the early changing point of the Lombard effect. Moreover, the results of Tang et al.12 are based on prediction models and the authors pointed out that their model is limited to the fact that the raised voice threshold was chosen empirically and thus it may be specific to the present experimental conditions. After the changing-point, the Lombard slope found in the present study was 0.54 dB/dB. This compares well with the Lombard slopes of 0.2 to 1 dB/dB reported in the literature,4 and with the Lombard slope of 0.69 dB/dB reported by Hodgson et al.13 based on a prediction model.
Regarding the subjective measures, the change-point for disturbance was lower at a noise level equal to 52.2 dB(A). Participants started to become (more) disturbed by noise at a lower noise level than with the associated change-point of the Lombard effect. Similar noise levels [51.3 and 52.5 dB(A)] are also triggering a decrease in the willingness to spend time and money in a restaurant. In conclusion, to improve the acoustic environment of restaurants, background noise levels should be lower than 50–55 dB(A). This will minimize the vocal effort of patrons and the disturbance in their communication. Concurrently, this will increase business for the restaurant since patrons would be willing to spend more time and money to eat in a restaurant with a background noise lower than 50–55 dB(A).
A limitation of this study is the absence of other patrons in the sound booth. This configuration probably inhibited possible automatic mechanisms in speech regulation related to privacy. Another limitation is associated with the reading material. After the first presentation, the participants become familiar with the excerpt contents and this could have an effect on the participants' subjective responses. However, if such effect were present, it would have been randomly distributed over the different conditions due to the randomized presentation of the noise levels. Future studies could refine the ecological validity of the present results, performing measurements in real restaurants.
Acknowledgments
The author would like to thank Kelly Kost, Emilie Palacios, Alison Perlman, Angelica Wozniak, and the participants for their involvement.