Objective Assessment of Speech Intelligibility in Crowded Public Spaces

Abstract

The objective of this study was to obtain a normative database of speech intelligibility data for young normal-hearing listeners communicating in public spaces. A total of 174 listeners participated in an interactive speech intelligibility task that required four-person groups to conduct a live version of the Modified Rhyme Test in noisy public spaces. The public spaces tested included a college library, a college cafeteria, a casual dining restaurant during lunch hour, and a crowded bar during happy hour. At the start of each trial, one of the participants was randomly selected as the talker, and a tablet computer was used to prompt them to say a word aloud from the Modified Rhyme Test. Then, the other three participants were required to select this word from one of six rhyming alternatives displayed on three other tablet computers. The tablet computers were also used to record the SPL at each listener location during and after the interval where the target talker was speaking. These SPL measurements were used to estimate the signal-to-noise ratio (SNR) in each trial of the experiment. As expected, the results show that speech intelligibility decreases, response time increases, and perceived difficulty increases as the background noise level increases. There was also a systematic decrease in SNR with increasing background noise, with SNR decreasing 0.44 dB for every 1 dB increase in ambient noise level above 60 dB. Overall, the results of this study have demonstrated how low-cost tablet computer-based data collection systems can be used to collect live-talker speech intelligibility data in real-world environments. We believe these techniques could be adapted for use in future studies focused on obtaining ecologically valid assessments of the effects of age, hearing impairment, amplification, and other factors on speech intelligibility performance in real-world environments.

INTRODUCTION

At one point or another, almost everyone has experienced the difficulties associated with carrying out a conversation in a noisy public space like a restaurant, a bar, or a crowded party. Individuals in modern society tend to congregate in public spaces to share meals and to socialize with their friends and relatives. When these spaces become crowded, the background noise level increases and conversation becomes increasingly difficult. Indeed, many listeners would agree that some of the most difficult and frustrating listening conditions they encounter in everyday life occur when they are conversing with friends and relatives in very loud public spaces.

Hearing scientists have long recognized how difficult it is for hearing-impaired listeners to be able to understand speech in crowded public spaces containing many competing talkers. Substantial effort has gone into understanding the challenges imposed by, and differences between speech and noise maskers (Miller 1947) and in developing hearing tests that use synthetic “cafeteria noise” and “babble noise” as an interfering sound source (Killion et al. 2004; Wilson et al. 2007). Additional efforts have been made to build complicated acoustic models for babble (Krishnamurthy & Hansen 2009) and to assess the differences between different types of babble and cafeteria noise on speech perception (Kaplan & Pickett 1982).

More recently, there has been a push to more closely examine the factors that influence speech perception in real-world public spaces. Some of these studies have been true “field” studies, where experimenters have used techniques like Ecological Momentary Assessment (EMA) to capture the background noise level and the subjective assessments of listeners who are conducting their everyday activities without modification in these real-world environments (Smeds et al. 2015; Wu et al. 2018; Holube et al. 2020). Others have been “hybrid” studies where auditory and/or audiovisual environments have been simulated through loudspeaker reproduction (Revit et al. 2002) or through sophisticated virtual audio and visual rendering techniques (Best et al. 2015; Hendrikse et al. 2019; Hohmann et al. 2020). However, until now, there have been relatively few hybrid studies carried out in the field that have objectively measured speech intelligibility performance in complex real-world environments. These types of studies are critical to our overarching goal of ensuring more ecologically valid outcomes, when the aim of research is to support the development of hearing-related procedures and interventions, as well as methods for assessing and predicting the ability of people and systems to accomplish specific real-world tasks, that is, purposes B (supporting the development of improved procedures and interventions) and C (facilitating improved methods for assessing and predicting the ability to accomplish real-world tasks) outlined in Keidser et al. (2020).

This is because such studies will allow us to ensure that our laboratory-based experiments are truly replicating the level of performance we would expect to obtain when talkers and listeners communicate with one another in the real world. For more in-depth discussions on various factors that play an important role in everyday communication situations, see Carlile and Keidser (2020), Grimm et al. (2020), and Lunner et al. (2020).

In this article, we describe the development of an experimental procedure capable of estimating signal-to-noise ratios (SNRs) and measuring the speech intelligibility performance of live talkers and listeners in real-world environments. We also present normative data from young, normal-hearing listeners in a variety of listening environments, and discuss the implications of these results and possible uses of the technique to address future research questions.

FACTORS INFLUENCING THE SNR IN REAL-WORLD PUBLIC SPACES

One of the most important factors influencing the speech intelligibility obtained in any environment is the SNR, which is generally defined as the ratio of the root mean-square (RMS) power of the target speech signal to the RMS power of the background noise masker at the location of the listener. By definition, the SNR is determined by two factors: the level of the target talker’s voice and the level of the background masking noise.

In real-world conversations, a theoretical argument could be made that the SNR value is almost entirely under the control of the talker. In some cases, some improvement in intelligibility may be possible simply by moving closer to the listener or leaning in to talk in the listener’s ear. If this is not sufficient, most talkers have enough dynamic range to overcome all but the highest levels of background noise. Most are capable of increasing their voices to levels in excess of 90 dB SPL at a distance of 1 m (Pearsons et al. 1977). Thus, depending on the average spectrum of the talker and the background noise, it is possible for a motivated talker to obtain an SNR of +10 dB or more simply by increasing the level of the voice. In reality, this does not happen, because talkers are subject to constraints that prevent them from talking loudly enough to be understood in any background noise environment. When they are in a public space, they may be under social pressure not to annoy others by speaking louder than is required to conduct a polite conversation. They may also be under some pressure to maintain speech privacy so other people in the space will not be able to eavesdrop on the substance of their conversation. And they may simply be constrained by vocal fatigue from using a raised voice level for an extended period (Beechey et al. 2018).

In quiet environments, talkers are generally adept at selecting a voice level that is loud enough to result in adequate speech intelligibility for a normal-hearing listener, but not too loud to comply with appropriate social norms for the surrounding space. Indeed, there are many popular culture references to the stigma associated with “loud talkers” and “quiet talkers” who habitually speak at levels that are louder or quieter than is considered appropriate for the situation (Hass 2009). Note, however, that there are some abnormal physiological conditions (e.g., Parkinson’s disease) that cause talkers to have difficulty achieving an appropriate speech level (Ho et al. 1999). Additionally, there are many cases where normal-hearing talkers in relatively quiet environments (public or private) speak too softly to be heard by individuals who have a hearing loss.

Furthermore, hearing-impaired talkers have difficulty assessing the loudness of their own voice when using hearing aids due to amplification, occlusion effects, and other factors (Jenstad et al. 2003; Laugesen et al. 2009).

In sparsely populated public spaces, the background noise is often dominated by general environmental noise (ventilation noise, clinking and clacking from a busy kitchen, background music) that is independent of the speaking level of a talker in the space.

However, as the number of individuals in a public environment increases, the contribution that conversational noise makes to the general background noise will start to increase.

Indeed, once the background noise reaches a certain level, the talkers will begin to involuntarily increase their own voice levels to overcome the noise and increase the intelligibility of their speech for the listener; this is known as the Lombard Effect (Dreher & O’Neill 1957; Junqua & Anglade 1990). Eventually, the background noise will become completely dominated by speech noise generated by other individuals who are having conversations in the same environment. Once this happens, the level of the background noise can no longer be viewed as independent of the speaking level of an individual talker in the space.

To a first approximation, all of the talkers in the environment can be viewed as experiencing the same challenge, where they are attempting to talk just loud enough to be heard over the background noise generated by the other talkers in the room. In this situation, individual talkers in the room can no longer obtain a lasting increase in SNR by speaking louder, because when they increase the level of their voice it will increase the background noise level for other talkers in the room, who in turn will respond by increasing their own speaking levels. Thus, it can be argued that, in these talker-dominated noise environments, the overall noise level effectively settles into an equilibrium where most talkers are speaking at roughly the same level and the SNR is roughly constant for all listeners in the room. The exact level of this equilibrium will depend on a number of acoustic factors, including (a) the size of the room; (b) the number of talkers in the room; (c) the average spacing between the listeners and talkers; (d) the reflectance of the physical surfaces and reverberation time of the room; (e) the presence of a loud acoustic masker, like amplified music; and, possibly, (f) the social environment and level of inhibition of the talkers.

It is important to note that the level of this equilibrium background noise level is not inherently stable. Any individual talker in the room could choose to increase their voice level enough to ensure 100% intelligibility for their listeners, and if they did so, the other talkers in the room would almost certainly react by raising their voices to compensate. If maximizing intelligibility were the only constraint, this would lead to a very unpleasant situation with all the talkers in the room shouting as loudly as possible. In practice, this rarely happens, because social pressures, fatigue, or perhaps even laziness or apathy conspire to cause talkers to choose a speaking level that is considered adequate for the situation even if it does not produce the highest achievable level of intelligibility. Indeed, even when they are talking loudly, talkers in these spaces may not be aware of their own voice level. The Lombard effect produces involuntary increases in speech level that are not apparent to the talker, and we adjust our own expectations for voice level based on the level of background noise. This phenomenon becomes very apparent in cases when someone is talking in loud noise and the background noise suddenly turns off: what seems like a normal conversational voice level in 80 dB of background noise is likely to seem like a loud shout in a quiet room.

Although most hearing researchers would agree that the SNR is the most critical determinant of speech intelligibility in noisy environments, there is surprisingly little information available about the actual SNR that listeners experience in real-world public spaces. In part, this is because the SNR in public spaces is difficult to measure. The overall background noise in a public space is easy to analyze, but it generally is not possible to get an isolated measurement of the level of the speech signal at the location of the listener. In general, studies that have attempted to measure SNR in real-world spaces have used mathematical techniques to estimate the SNR by comparing the overall level of a representative audio sample containing both the target speech and background noise and comparing it to the overall level of a representative sample containing only the background noise (Pearsons et al. 1977; Smeds et al. 2015; Wu et al. 2018). The “noise+target” and “noise alone” samples are often manually identified by an experimenter listening to the samples a posteriori. This procedure can produce an approximation of the SNR, but it is not able to provide a precise measure of SNR, particularly in cases where the SNR is close to or less than 0 dB.

One question that has substantial practical importance but has not been extensively addressed in prior studies is the actual level of speech-in-noise performance one can expect listeners to achieve in real-world environments. Communication is a two-way process, and one can imagine that at least part of what determines the speech level and thus the SNR in a real world environment is the talker’s use of visual and nonverbal cues to assess how well a listener is able to understand them in that particular environment. Visual speech cues can also be very important for listener performance (Sumby & Pollack 1954). These visual cues are almost always available in the real world but are only occasionally provided in laboratory studies of speech perception. Listeners in real-world environments also have access to information about the physical configuration of the room and the noise sources within it that would typically not be available to a listener hearing exactly the same acoustic stimulus in a laboratory environment. In extreme environments, talkers and listeners may be able to develop joint strategies (like leaning in to talk to a listener’s ear) that would be difficult to replicate in a laboratory environment (Hadley et al. 2019). Both talkers and listeners may also be exposed to visual and auditory distractions in real-world environments that would be hard to capture in the laboratory.

The goal of this study was to evaluate how well groups of listeners, drawn from a large population of young normal-hearing listeners, were able to understand each other when they were speaking and listening to one another in a variety of quiet and noisy public spaces. A secondary goal was to develop a low-cost, highly portable measurement system that could be used for future experiments focused on speech intelligibility in real-world environments. The next section describes the methods used in more detail.

METHODS

Participants

The data were collected in a series of experiments conducted at the University of Maryland between 2014 and 2019 under a research protocol approved by the University of Maryland (UMD) Internal Review Board. Speech intelligibility data were analyzed for a total of 174 participants, all of whom served both as talkers and listeners in the experiment. The listeners were all young (18–35 years, mean age 21.3 years) and all had normal hearing, which was confirmed by screening each participant’s hearing at 20 dB HL (re: ANSI 2018) at octave frequencies between 250 and 8000 Hz using an Interacoustics AC-40 audiometer with ER-3A insert earphones in a sound-attenuated booth. All of the participants identified themselves as either native or bilingual speakers of American English.

Most of the participants were college undergraduates and were compensated with class credit or a small monetary payment. However, six of the participants were experimenters in the study who filled in for one of the listeners when there were not enough participants at a test session to make up a full four-person group. These six participants accounted for 8.9% of the total trials in the experiment, with no single individual accounting for more than 3.3% of the trials.

Stimuli

The Modified Rhyme Test (MRT) (House et al. 1965) was used to measure speech intelligibility in the study. The MRT consists of 274 monosyllabic consonant–vowel–consonant (CVC) words, which are presented in a closed-set format where the listener chooses between the target word and five rhyming foils that differ from the target word only by either the initial or final consonant phoneme. The target word was presented live voice by the designated speaker (paradigm described below) in the carrier phrase, “You will mark (target) please.” The MRT was well suited to this study, because its closed set format allowed automatic scoring of the listener responses on the tablets. Although it uses a very simple set of speech stimuli, the MRT is known to be related to overall performance in critical communication tasks (Letowski & Scharine 2017), and it is one of the standard live-talker tests included in the ANSI S3.2 standard for measuring speech intelligibility over communication systems (ANSI, 2009). Its rhyme-based format also allows it to be used to evaluate confusion rates between different individual phonemes.

Test Locations and Noise

Participants completed the task in at least one of five test locations: a quiet library, a dining hall at UMD, a local restaurant, a pool hall at the student union at UMD, and a local bar. In each listening environment, the noise consisted of the natural background noise that was present in the chosen environment during the testing, as described in the Procedure section. The noise varied in level from moment to moment, and generally included a mix of talkers, furniture movement, voices from a television, and/or clatter of dishes and barware. An example of each listening environment can be seen in Figure 1. Descriptions of the listening environments of the five test locations were as follows:

Fig. 1.

Pictures of four environments that were used for testing in the experiment. (A) Library; (B) Cafeteria; (C) Restaurant; and (D) Bar.

• Quiet Library: The quiet library environment had a large open-space design, with several adjacent tables. The tables were approximately 8–10 feet apart, and there were typically 1–2 quiet patrons at each of the adjacent tables.

• Cafeteria: The cafeteria was a dining hall at UMD with a large open-space design containing approximately 50 tables. Adjacent tables were approximately 3–4 feet apart, and typically seated 3 or more patrons who talked at normal conversational levels. The dining hall also had several televisions mounted throughout the cafeteria. Participants were typically seated at a table close to one of the televisions. The dining hall had high ceilings and tiled flooring.

• Restaurant: The restaurant environment was a casual dining restaurant from a national chain that was situated near the university. The restaurant used a large open-space design, with about 15 tables situated approximately 3–4 feet apart. There were typically 2 or more patrons at each table who talked at normal conversational levels. The kitchen area was open to the dining area, and participants were seated at a table closest to the kitchen to optimize the amount of ambient noise. The restaurant was not sound treated. It had tile flooring and metal fixtures completely covering all of the walls and ceilings.

• Pool Hall: The Pool Hall environment was an on-campus multipurpose gaming hall with eight bowling lanes, 10 billiards tables, two projection screen televisions, and several other televisions mounted throughout the hall. Adjacent tables were approximately 3–4 feet apart, and typically seated three or more patrons who talked at normal conversational levels. The hall had a drop ceiling, and some portions of the floor were tiled, other portions of the floor were carpeted.

• Bar: The bar environment was a large commercial pub with a popular happy hour catering to college-age patrons. It had an open-space design with approximately 20 tables for dining, a separate bar area, and a stage for music performances. Adjacent tables were situated approximately 3–4 feet apart, and typically seated two or more patrons who talked at normal to louder-than-normal conversational levels. The bar environment had numerous televisions mounted throughout the bar. The participants were typically seated at a table close to mounted televisions and near the bar, in order to optimize the amount of ambient noise from fellow patrons, televisions, and barware. The bar had very high ceilings, hard wood floors, and large windows.

Data Collection

Groups of four participants were recruited to participate in 71 listening sessions in the study. The sessions were conducted in different locations, and each session was further divided into blocks of trials that typically included 52 trials. Table 1 shows the distribution of blocks, subjects, and trials across the five locations tested in the experiment. Overall, a total of 12,768 trials were collected in the experiment, divided across 195 blocks of trials.

TABLE 1.

Number of blocks and listeners collected at each location in the experiment

Location	Listeners	Blocks	Trials
Library	17	17	720
Cafeteria	37	50	2,609
Restaurant	86 (84)	108 (107)	6,023 (5,989)
Pool hall	10 (4)	5(3)	517 (393)
Bar	34 (7)	15 (8)	2,889 (762)
Total	174 (132)	195 (185)	12,768 (10,473)

The numbers in parentheses show the number of conditions where SNR data were available (where they differed from the overall total). The number of trials represents the number of standard MRT trials collected in that condition; in some cases, additional trials may have been collected in other listening tests that are outside the scope of this manuscript. Note that some listeners participated in more than one environment, so the total number of listeners is less than the sum of the first column.

Procedures

Test locations and time of testing were chosen to maximize ambient noise level. If participants failed to appear for the experiment, canceled, or if a complete group of four participants could not be recruited in time for testing, one to two experimenters would sit in as participants in the experiment. This occurred for 27 separate data collection sessions. During the experiment, the four members of the group were seated at a rectangular table, with two participants seated on each long-side of the table. The average distance between participants, measured at the location of the heads, was approximately 1 m. A handheld tablet computer was placed on a stand on the table in front of each listener. This tablet was used to provide instructions to the participants and to allow them to make their responses. At the start of each trial, one participant was randomly selected to serve as the talker, and a prompt appeared on that participant’s screen telling them to read the carrier phrase and target word out loud when it appeared on his or her tablet. After saying the carrier phrase, the talker was instructed to tap the tablet screen to indicate they were done talking. This caused the six response words to appear on screens of the three other participants, who served as listeners. The three listeners responded by tapping the appropriate target word on the screen. The tablet then recorded the participants’ responses during each trial. The tablets also recorded the response time of the listeners in response to the spoken carrier phrase. The response time was the interval between when the target talker tapped their tablet to trigger the response matrix for the listeners (after reciting the carrier phrase) and the listener’s tapping response on their individual tablet. Individuals were instructed to stay seated in their initial positions, and they were told to speak as they normally would in the environment where the testing occurred.

Note that, because of the design of the experiment, there was no way to know precisely how long it took the target talker to start talking after they were first given the prompt to start, or how long they waited to tap the screen after they were done. We only know that the target speech occurred at some point during the talker interval, and that there was no target speech in the response interval. This is why it is necessary to use the peak level in the target interval, rather than the mean level, to estimate SNR. The ambiguity about when the speech actually occurred in the target interval also made it necessary to measure listener response time from the start of the response interval (when they were first shown the six alternative MRT responses), rather than from the actual time when the target talker said the key word.

Most of the sessions used Nexus 7 tablets that were connected individually to a Dell laptop personal computer via Bluetooth fireplug USBs. As the individuals identified the target word on their tablet, the accuracy and response time data were uploaded to the laptop PC using a code written in MATLAB (Mathworks, Natick, MA). In some testing sessions, a Larson–Davis System 824 sound level meter (SLM) was placed in the middle of the table to measure the SPL of the background noise and target speech during each trial. The individual participant tablets also measured the SPL at the location of each participant. The final five sessions used a new setup where four Samsung Tab-E tablets were connected to a fifth control tablet, and sound level measurements were made with a Svantek SV104 SLM.

Participants were tested for 2 hours over the course of a single session, with one 10–15 minute break at approximately 1 hour. The participants performed as many blocks of the experiment as they could as a group within the two-hour time frame. Each block typically consisted of 52 trials. However, some of these trials were in conditions other than the baseline MRT task discussed in this paper, and the number of trials per block increased in some of the later sessions of the experiment, so the number of actual MRT trials per block could vary from less than 10 to more than 100. On average, participants completed roughly 200 MRT trials in a single session.

After completing each block of trials, each listener rated each talker, based on the perceived difficulty of understanding that particular talker. A Likert scale with rating categories from 1 to 7 was utilized for the ratings. A rating of “1” signified that the listener had no difficulties at all understanding the talker, and “7” indicated that the listener could not understand the talker at all. Background noise level, accuracy (percent correct score), log response time, and rating of talker difficulty were measured for each listener for each block.

Trial Sequence

Before the start of each trial, one of the four participants was randomly selected to serve as the target talker. Then, at the start of the trial, the target talker’s screen changed to a prompt telling the talker to say an MRT carrier phrase of the form “You will mark BAN please” and then tap the screen. At that point, the target talker’s screen became blank and the other participants were shown a list of six MRT response words. The response interval then extended until the last of the three listeners made their response.

Figure 2 is a timeline showing auditory and response data collected during a single trial of the experiment. The trial time started at the point when the target talker was first prompted to say the carrier phrase. Then the response interval started at the point when the talker tapped the screen to indicate they were done speaking, and the six response words were shown to each of the three listeners. The transition between talker and response interval is shown by the vertical black line in the figure. The vertical dotted lines in the response interval show the response times of the three listener responses. The response interval ended when the last of the three listeners made their response.

Fig. 2.

Example of sound levels recorded at a single microphone level during a typical trial in the experiment. The talker interval extends from the time when the talker is first prompted to say the MRT carrier phrase (t = 0 s) until the time when the talker taps the screen (t = 3 s) to indicate they are done speaking. The vertical dashed lines show the response times of the three different listeners. The response interval ends when the last listener responds by sg a reelectinsponse MRT word (t = 7 s). MRT, Modified Rhyme Test.

The blue line in the figure shows the overall dBA level measured at one of the four on-board tablet microphones used in the experiment. The onboard sound level monitoring software on the tablets continuously monitored the SPL in a fast averaging mode (with a 125 ms time-constant) and these ongoing sound-level measurements were recorded every 0.5 s, starting at the beginning of the trial.

Three different noise values were calculated for each microphone location in each trial of the experiment. The light blue circle shows SPL Peak_Talk, which is the peak dBA SPL measured during the talker interval of the experiment. The dark blue square shows SPL Peak_Resp, which is the peak dBA SPL measured during the response interval of the experiment. And the horizontal dashed line shows the background noise level (SPLBack), which was calculated from the mean dBA level averaged across the entire response interval.

The major data points collected in each trial were:

1. The peak sound level in the talker interval, SPL Peak_Talk

2. The peak sound level in the response interval, SPL PeakResp

3. The average sound level in the response interval, SPLBack

4. The responses and response times of the three listeners

5. The subjective talker difficulty ratings (which were made at the end of each block)

RESULTS

Noise Level Measurements

Reliability of Tablet-Based Sound Level Measurements •

One key aspect of the experimental design of this experiment was the use of low-cost, lightweight Android tablet computers as a way of measuring sound levels and collecting subject responses. However, although there are some data on the use of Android and iOS portable devices to measure ambient sound levels, there are still relatively little data available on the reliability of these measurements. In this experiment, we initially included a commercial Type-I SLM (Larson Davis 824) that was placed in the center of the table where the four participants were seated. This SLM was set in fast averaging mode (125 ms time-constant), and, as was the case with the tablet measurements, the experiment software was set up to record this SPL every 0.5 sec from the start of the trial.

Although the location of the commercial SLM differed somewhat from the locations of the four built-in microphones on the tablets, it is still possible to compare the tablet and SLM measurements to obtain an estimate of the reliability of the tablet-based microphones.

A total of 73 blocks of trials (totaling 3,235 individual trials) were collected with sound level measurements from both the commercial SLM and the on-board tablet microphones. To account for differences between the individual tablets, four different calibration factors (constant dB offsets) were determined to adjust the mean SPLBack value measured with the on-tablet microphone to match the mean SPLBack value measured with the commercial SLM. These individual calibration factors ranged from +2.7 dB to −0.5 dB. After applying these calibration adjustments, the standard deviation of the error between the tablet SPL measurements and the commercial SLM measurements was 1.7 dB. The correlation coefficient between the tablet and SLM measurements was 0.85. This level of reliability was determined to be adequate for the purposes of the experiment, particularly when one considers that some of this variation represents true variations in the actual sound level at the four tablet locations. On the basis of these data, a decision was made to discontinue the use of the commercial SLM and rely on the tablet-based sound measurements. All subsequent analyses in this article will be based on the tablet-based sound measurements.

Distribution of Peak and Background Noise Levels

Figure 3 shows the distributions of peak and average sound levels measured during the talker and response intervals of the experiment. The different symbols in the figure represent the six different types of noise measurements (as described below).

Fig. 3.

Average values were obtained for 6 different types of noise measurements that captured the peak and average SPL levels recorded by the tablet microphones during each trial of the experiment. The data have been separated into five sections according to the environments where they were collected. Within each noise environment, the blocks have been sorted in order of increasing the background noise level. See text for details.

Note that the blocks of trials have been divided into the five different noise environments tested in the experiment, and that within each environment the blocks have been sorted to have increasing SPLBack levels moving from left to right in the figure. The six noise values shown are:

• Peak_Talk (Talker), which is the peak SPL during the talker interval of the trial measured at the talker’s tablet;

• Peak_Talk (Listener), which is the peak SPL during the talker interval of the trial measured at one of the listener’s tablets;

• PeakResp (Talker), which is the peak SPL during the response interval of the trial measured at the talker’s tablet;

• PeakResp (Listener), which is the peak SPL during the response interval of the trial measured at one of the listener’s tablets;

• Mean SPLBack (Talker), which is the mean SPL during the response interval of the trial measured at the talker’s tablet;

• Mean SPLBack (Listener), which is the mean SPL during the response interval of the trial measured at one of the listener’s tablets;

These results show the following important trends:

1. In the talker interval of the experiment, the peak sound level at the talker’s tablet was systematically higher than the peak sound level at the listeners’ tablets. This can be seen by comparing yellow symbols to the purple symbols in each panel of the figure. Averaged across all the blocks of the experiment, the value of Peak_Talk at the talker tablet was 2.7 dB higher than the value of Peak_Talk measured at the listener tablet (σ = 1.2 dB). This is consistent with the geometric configuration used in our study, where the tablet located on the table in front of the talker would typically be 30%–40% closer to the target talker than the tablets located on the table in front of the listeners. The correlation coefficient between the peak levels at the talker and listener conditions (averaged across blocks) was 0.97.

2. There was essentially no difference between the peak or mean sound levels measured at the talker or listener tablets during the response interval. This can be observed by noting that the dark blue and red symbols were nearly identical in all blocks, and that the light blue and green symbols were also nearly identical in all blocks. This is an extremely important finding, because it confirms that the sound measurements made during the response interval were all measuring the ambient noise level in the environment, which one would expect to be (on average) the same at the locations of all four participants. If some portion of the target talker’s speech was contained in the response interval, one would expect a systematic trend for the sound level to be higher at the target talker’s location than at the listener locations (as it was for the measurements made in the talker interval of the experiment).

3. The ratio between the peak SPL measured in the response interval and the average SPL measured in the response interval was very consistent in the louder public spaces. Across all four environments, the mean value was 1.7 dB (σ = 0.89 dB). This reflects the fact that the background noise in the public spaces tended to be relatively stable in terms of overall level, as one would expect for a multitalker babble containing many competing talkers.

Estimated SNR as a Function of Background Noise Level

In this experiment, the masker-only and target-plus-masker intervals were automatically separated by having the target talker terminate the talker interval by indicating when they were finished speaking. However, as noted before, there is no way to know where within the talker interval the target speech actually occurred. Although it is not possible to calculate the true SNR value from the relatively crude sound level measurements made in this study, it is possible to estimate a proxy for the true SNR by taking the difference between the peak sound level in the talker interval Peak_Talk and the peak sound level in the response interval PeakResp. This estimate, which we refer to as SNRP/P, is based on the assumption that the target talker will be the loudest sound source during the target interval, and consequently that the peak SPL that occurs in the talker interval is equivalent to the peak SPL of the target speech. If the response interval only contained a single speech signal, then the SNRP/P would be an unbiased estimate of the true SNR. However, the response interval generally contained a mixture of many talkers, which means that the peak-to-mean ratio of the background noise contained in the response interval is likely smaller than the peak-to-mean ratio of the target speech signal. This means that the SNRP/P is likely an overestimate of the true SNR, which is traditionally calculated as the ratio of the overall RMS levels of the target and masking waveforms.

To obtain an estimate of the true SNR, a simulation was conducted where one of the tablets was placed on a stand in the center of a 10-loudspeaker array. Nine of those speakers were used to simulate nine different talkers producing continuous sentences in an arc around the tablet. The 10th loudspeaker, which was located directly behind the tablet (where a target talker might be expected to be located in this study) was used to simulate the target speech. The level of the target talker relative to the nine masking talkers was systematically adjusted, and the SNRP/P (which was calculated by comparing the peak level encountered over a 3-sec interval with all 10 talkers to the peak level encountered over a 3-sec interval with just the masking talkers) was measured as a function of the true SNR (which was calculated by comparing the level of the target talker at the tablet location with no maskers active to the level of the masking talkers at the location of the tablet with no target talker). Note that these 3 sec long intervals were randomly sampled 20 sec long recordings of the simulated talker and response intervals that were made at each SNR value. The results showed that the SNRP/P was almost exactly 5 dB higher than the true SNR over the entire range of SNRs measured in this experiment. Thus, we can approximate the true SNR of the speech signal simply by subtracting 5 dB from the SNRP/P.

Figure 4 shows the mean SNR_P/P (left axis) and estimated true SNR (right axis) as a function of the mean SPLBack for the experimental blocks that contained at least 10 trials. The different symbols represent the five different listening environments where SNR estimates were available. These results indicate that there was a systematic, roughly linear relationship between the SNR and the SPLBack when the background noise exceeded 60 dBA. The upper left hand of the figure shows the equation for a linear fit for SNR_P/P as a function of SPLBack for all SPLBack values greater than 60 dBA. This equation reveals that the SNR decreased by 0.44 dB for each 1 dB increase in the SPLBack. The correlation coefficient between individual trials with an SPL level greater than 60 dBA and the corresponding proportion of correct responses was 0.56. The correlation coefficient between individual blocks of trials with a mean SPL level greater than 60 dBA and the corresponding mean proportion of correct responses was 0.85.

Fig. 4.

Mean estimates of SNR _P/P as a function of SPL_Back for blocks of the experiment that contained data from more than 10 trials. The dotted line, equation, and correlation coefficients correspond to the linear regression curve between the SNR _P/P and SPL_Back values for trials where the SPL_Back was greater than 60 dBA SPL. The gray line shows the approximation of SNR vs. background noise level from Wu et al. (2018). SNR, signal-to-noise ratio.

This equation was relatively consistent for the Cafeteria, Restaurant, and Bar locations, but did not appear to hold for the Library environment, where the background noise level was generally less than 60 dBA SPL. Note that the gray line on the figure represents the estimate of SNR as a function of SPLBack reported by Wu et al. (2018). The estimate provided by Wu and colleagues very closely matches the results of this experiment, with the exception of the region around 65 dB. The data are also similar to those reported by Weisser and Buchholz (2019), who measured speech levels in a laboratory setting where two individuals completed a face-to-face communication task while listening to simulated background noise presented over acoustically open headphones. They found that the SNR decreased by 0.56 dB per 1 dB increase in background noise, compared to 0.44 dB/dB in this study. After controlling for the adjustments in distance that talkers and listeners tend to make in noisy environments (Pearsons et al. 1977), they found that the SNR would generally be expected to fall below 0 dB when the background noise level exceeds 75 dBA, which is essentially identical to the value we obtained for the 0 dB SNR intercept in our data (75.04 dB). However, those estimates would predict that individuals would reduce their interpersonal distance to 0.5 m when the noise exceeded 75 dBA, which is closer than the roughly 1-m spacing between talkers and listeners that occurred in this experiment.

Performance as a Function of SNR

The left panel of Figure 5 shows the proportion of correctly identified words as a function of the estimated SNR for each block of the experiment that had valid data for at least 10 trials in the MRT task. From this graph, it is apparent that there is a general trend for performance to decrease as the SNR decreases, but that this trend is not consistent across the different environments. In particular, it is clear that performance in the Library environment was consistently near perfect even when the measured SNR_P/P was near 0 dB.

Fig. 5.

The left panel shows performance as a function of the estimated true SNR for individual blocks containing more than 10 trials. The gray shaded region shows the 95% confidence interval of a logistic on the data. Blocks that have been plotted over a red ‘*’ were collected in a background noise level less than 60 dB SPL. These blocks, which were primarily obtained in the Library condition, exhibited high performance regardless of estimated SNR and were fitted with a separate logistic regression, which is shown by the black dotted line. The right panel shows performance as a function of background noise for individual blocks containing more than 10 trials, along with a shaded region showing the 95% confidence interval of the logistic regression. SNR, signal-to-noise ratio.

In part, the differences in performance between the different listening environments appear to be explained simply from the lower overall noise level that occurred in the library and cafeteria environments. The symbols in Figure 5 that are plotted over a red asterisk are trials where the background noise was less than 60 dBA. A logistic fit to these trials (dotted line) shows that listeners scored about 95% correct on these trials, regardless of SNR.

The other symbols show mean percent correct as a function of SNR for trials where the background noise level exceeded 60 dBA. The solid black line and surrounding shaded area shows the 95% confidence interval for a logistic fit to these trials. This logistic estimate shows that the estimated percent correct increases systematically with SNR, from about 82% when the SNR was −5 dB to 95% when the SNR was +10 dB.

Performance as a Function of Background Noise Level

The right panel of Figure 5 shows performance in the MRT task as a function of the background noise level SPLBack for all the individual blocks of trials of the experiment with more than 10 trials of data. These results show a clear, systematic effect of background noise on overall performance in the MRT task. The shaded region of

Figure 5 shows the 95% confidence interval of a logistical fit to all the individual data points collected in the experiment. Although there was substantial variation across blocks that had the same average background noise level, these data suggest there is a clear systematic relationship between the overall noise level in a public space and the average intelligibility level one would expect young normal-hearing listeners to obtain in that environment. When the noise level exceeds 85 dBA SPL, average intelligibility falls to less than 80% on the MRT.

Response Time

Another measure of hearing difficulty in a speech intelligibility task is the response time listeners require to identify the target speech signal. In a typical speech response time task, the response time starts at the point when the target speech signal is presented and ends when the listener makes a response with a button, mouse, or touch screen. In this study, it was impossible to know the exact time the speech signal was presented within the talker interval. Thus, the response time had to be measured from the time when the six alternatives were first presented on the screens of the listeners (which occurred when the talker pressed the “done” button to end the talker interval) up until the time when the listener selected one of those six alternative responses. The left panel of Figure 6 shows the response time in the experiment as a function of the background noise SPLBack. These results show there was a marginally significant correlation between response time and background noise level, with response time increasing by roughly 27 ms per dB of background noise (p = 0.08). Response times for SPL levels greater than 80 dB were significantly slower than those for SPL levels less than 70 dB (two-tailed t-test, p = 0.0001).

Fig. 6.

Log response time as a function of SPL_Back(left panel) and as a function of percent correct (right panel). In order to reduce the impact of spurious responses, the slowest 5% and fastest 5% of response times in each block of trials were eliminated from this analysis.

A significantly stronger correlation was found between the mean response time and the percentage of correct responses in each block of trials (fisher r-to-z transformation, p < 0.001). This is shown in the right panel of Figure 6.

Difficulty Ratings

The last measure of performance in the different environments was the subjective rating. The left and right panels of Figure 7 show the mean self-reported difficulty ratings obtained in each block of the tasks as a function of the mean background noise SPLBack and the proportion of correct responses obtained in each block of the experiment. As would be expected, these results show that perceived difficulty increased with increasing background noise level and with decreasing performance.

Fig. 7.

Difficulty rating as a function of SPL_Back (left panel) and as a function of percent correct (right panel) of the individual blocks of the experiment. The small symbols show blocks with 10 to 50 trials, the medium symbols show blocks with 51 to 100 trials, and the large symbols show blocks with 100 or more trials.

DISCUSSION AND CONCLUSIONS

The overarching goal of this study was to evaluate how well young, normal-hearing listeners were able to perform in a Modified-Rhyme speech intelligibility task that was conducted in a variety of real-world public spaces. The major results of this study can be summarized as follows:

• At background noise levels greater than 60 dBA, both the effective SNR and the speech intelligibility of young, normal hearing listeners conversing in public spaces appears to systematically degrade. The SNR decreases by roughly 0.4 dB per dB of background noise above 60 dB, and speech intelligibility drops from ceiling performance (roughly 95%) to roughly 80% correct performance as the background noise increases from 60 dBA to 85 dBA. Note that 80% on the MRT test should not be considered a high score. As a point of reference, a score of 80% on the MRT is considered to be the minimum acceptable level for communications in the human factors standards in the U.S. Federal Aviation Administration and Department of Defence, and a mean MRT score in excess of 90% is required to meet the minimum intelligibility standard for public spaces in ISO 2470:2007 (Letowski & Scharine 2017).

• At noise levels below 60 dBA, normal-hearing listeners have no difficulty communicating in public spaces. Mean performance was always above 90% correct in these noise environments. This was true even in blocks where the estimated SNR value was very low. The reason for the decoupling of SNR and percent correct at levels below 60 dB is that this is the point where the background noise level in public spaces stops being dominated by the voices of other competing talkers and starts to be dominated by ventilation noise, kitchen noise, or other environmental noise sources. These discrete environmental noise events (coughing, footsteps, door slamming, etc.) had a relatively high peak level but occurred at a different time or in a different frequency range than the target speech and thus did not have a large influence on the actual SNR that occurred when the target talker was talking.

• Response time in the MRT task systematically increased with increasing noise levels and decreasing percent correct scores. These response time effects were fairly weak, but they are notable because they are present even in an across-subject design where entirely different groups of talkers and listeners were tested in the different listening environments.

• Perceived difficulty, as measured by a subjective rating scale, systematically increased with increasing background noise, decreasing SNR, and decreasing percent correct performance.

• A low-cost data collection system based on Android tablets does appear to be capable of reliably collecting data on stimulus levels and speech intelligibility performance in real-world environments. After calibration, the background noise measurements made by the tablets were comparable to those obtained with a high-quality commercial SLM. Also, the SNR estimates that were automatically collected by measuring peak SNR level during the talker and response intervals generated SNR estimates comparable to those measured in earlier studies with different techniques (Pearsons et al. 1977; Smeds et al. 2015; Wu et al. 2018).

One notable difference between this study and the earlier studies of SNR in real-world environments by Wu et al. (2018) and Smeds et al. (2015) is the inclusion of listening environments with a much higher level of background noise. These prior two studies focused on listeners with hearing impairment and attempted to estimate the distribution of noise environments encountered by these listeners as part of their everyday lives. The results indicated that these listeners spent only a very small proportion of their time in environments where the SNR value was less than 0 dB (4%–15%). The current study was more like the Pearsons et al. (1977) study, where individuals were intentionally recruited to go to noisy public spaces, without regard to their proclivities to spend time in these environments in their everyday lives. These noise environments included spaces that were definitely within the normal range of experiences for young adults (a cafeteria, a popular casual food restaurant, a bar at happy hour), but we have no way of determining what percentage of time they would normally spend in these spaces. However, it is not unreasonable to think that all listeners might bias their overall perception of hearing difficulty to focus on the most difficult situations (similar to those tested here) even if they only spend a very small portion of their lives in these environments.

In closing, it seems appropriate to briefly discuss this experiment in the context of the framework for evaluating ecological validity that was discussed at the beginning of this supplement (Keidser et al. 2020). In that framework, this article would be considered a hybrid study, because the experimenters had tight experimental control over some variables, like the specific words spoken by the target talker on each trial, but almost no experimental control over other variables, like the auditory and visual distractions that might occur in the background during the test session. Clearly, the primary focus of this study was to ensure a high level of the “Environment” factor of ecological validity. The public spaces where the study was conducted were the same kinds of public spaces where the participants and their peers go everyday to engage in group conversations. No compromises or assumptions had to be made to ensure the realism of the acoustic field, either in terms of the levels, locations, and types of acoustic maskers or in the acoustical property of the room. In terms of the “Sources of Stimuli” factor of ecological validity, this study scored relatively high: the stimuli were realistic in the sense that they were generated by a variety of different live talkers who had a range of different voice characteristics, speaking styles, and levels of motivation to speak clearly and at an appropriate level for the acoustic environment. However, the MRT carrier phrases used as the speech materials were very artificial. They lacked any contextual cues, and would not have stimulated any level of emotional engagement from any of the participants. The study would also score somewhat lower in terms of “Task.” The MRT task of identifying rhyming words is not something anyone would really ever do in a noisy environment (although there is ample evidence that scores on the MRT are predictive of intelligibility for more complex speech materials). Indeed, the MRT is a popular task in laboratory speech intelligibility testing because it minimizes differences across talkers and listeners and, in particular, it limits the effects of cognition and learning on overall performance (Letowski & Scharine 2017). In the real world, cognitive processing is likely to play a large role in how successful a listener will be at conversing in a noisy space. In terms of “Context of Participation,” the present study is somewhat mixed. Although the MRT task itself is unlikely to motivate participants beyond what would be normally expected for paid volunteers in psychophysical experiments, the rotation of the role of talker across the participants in this study probably did cause individuals to invest more effort into doing well on the task than they might if they were sitting alone in front of a computer monitor while wearing headphones. Finally, in terms of “individual” ecological validity, this particular study is limited in that most of the participants were young normal-hearing listeners, and for the moment, our results are really only valid for that group. However, we believe there is great promise for extending this paradigm to explore how individual differences in factors like age, hearing impairment, and language proficiency influence intelligibility in real-world environments, as well as the impact that the use of hearing aids or personal protective equipment might have on real-world speech production and perception. Studies are now underway to explore these factors in more detail.

ACKNOWLEDGMENTS

We would like to thank Kerrianne Clark and Mikayla Abrams for their assistance in collecting data for this study. The identification of specific products or scientific instrumentation is considered an integral part of the scientific endeavor and does not constitute endorsement or implied endorsement on the part of the author, DoD, or any component agency. The views expressed in this manuscript are those of the authors and do not reflect the official policy of the Department of Army/Navy/Air Force, Department of Defense, or the U.S. Government.