Characteristics Of Real-World Signal To Noise Ratios And Speech .

294WU ET AL. / EAR & HEARING, VOL. 39, NO. 2, 293–304activity categories (e.g., conversation in a group more than threepeople) and five environmental categories (e.g., moving traffic), resulting in 30 unique listening situations. The frequencyof occurrence of each listening situation and the mean overallsound level of several frequent situations were reported.More recently, Wolters et al. (2016) developed a commonsound scenarios framework using the data from the literature.Specifically, information regarding the listener’s intention andtask, as well as the frequency of occurrence, importance, andlistening difficulty of the listening situation, was extracted orestimated from previous research. Fourteen scenarios, whichare grouped into three intention categories (speech communication, focused listening, and nonspecific listening), weredeveloped.Speech Listening and Signal to Noise RatioAmong all types of listening situations, it is arguable thatspeech listening is the most important. Although previousresearch (Jensen & Nielsen 2005; Wagener et al. 2008; Wu& Bentler 2012) reported the overall sound level of typicalreal-world listening environments, none provided informationregarding the signal to noise ratio (SNR) of speech listeningsituations. SNR is highly relevant to speech understanding(Plomp 1986) and has a strong effect on hearing aid outcome(Walden et al. 2005; Wu & Bentler 2010a). Historically, Pearsons et al. (1977) was one of the first studies to examine SNRsof real-world speech listening situations. In that study, audiowas recorded during face-to-face communication in variouslocations, including homes, public places, department stores,and trains using a microphone mounted near the ear on an eyeglass frame. Approximately 110 measurements were made. Foreach measurement, the speech level and SNR were estimated.The results indicated that when the noise level was below45 dBA, the speech level at the listener’s ear remained at a constant 55 dBA. As noise level increased, speech level increasedsystematically at a linear rate of 0.6 dB/dB. The SNR decreasedto 0 dB when the noise reached 70 dBA. Approximately 15.5%of the measurements had SNRs below 0 dB.The data reported by Pearsons et al. (1977) have beenwidely used to determine the SNR of speech-related tests forindividuals with normal hearing or with hearing loss. However,the participants in Pearsons et al. were adults with normal hearing. More recently, Smeds et al. (2015) estimated the SNRsof real-world environments encountered by hearing aid userswith moderate hearing loss using the audio recordings madeby Wagener et al. (2008). The speech level was estimated bysubtracting the power of the noise signal from the power of thespeech-plus-noise signal. A total of 72 pairs of SNRs (fromtwo ears) were derived. The results were not completely inline with those reported by Pearsons et al. (1977). Smeds et al.(2015) found that there were very few negative SNRs (approximately 4.2% and 13.7% for the better and worse SNR ears,respectively); most SNRs had positive values. At a given noiselevel, the SNRs estimated by Smeds et al. (2015) were 3 to 5dB higher than those reported by Pearsons et al. (1977), especially in situations with low-level noise. In quiet environments(median noise 41 dBA), the median speech level reported bySmeds et al. was 63 dBA, which was higher than that reportedby Pearsons et al. (55 dBA). Smeds et al. suggested that thediscrepancy between the two studies could be because of thedifference in research participants (hearing aid users versusnormal-hearing adults) and the ways that recordings were collected and analyzed.Visual Cues and Speech/Noise LocationOther than SNR, there are real-world factors that can affectspeech understanding and hearing aid outcome and that shouldbe considered in ecologically valid laboratory testing. Forexample, visual cues, such as lipreading, are often available inreal-world listening situations. Visual cues have a strong effecton speech recognition (Sumby & Pollack 1954) and have thepotential to influence hearing aid outcomes (Wu & Bentler2010a, b). Therefore, some speech recognition materials canbe presented in an audiovisual modality (e.g., the ConnectedSpeech Test; Cox et al. 1987a). Another example is the location of speech and noise sources. Because this factor can affectspeech understanding (e.g., Ahlstrom et al. 2009) and the benefit from hearing aid technologies (Ricketts 2000; Ahlstrom etal. 2009; Wu et al. 2013), researchers have tried to use realisticspeech/noise sound field configurations in laboratory testing.For example, in a study designed to examine the effect of asymmetric directional hearing aid fitting, Hornsby and Ricketts(2007) manipulated the location of speech (front or side) andnoise sources (surround or side) to simulate various real-worldspeech listening situations.Only a few studies have examined the availability of visualcues and speech/noise locations in real-world listening situations (Walden et al. 2004; Wu & Bentler 2010b). Wu andBentler (2010b) asked adults with hearing loss to describe thecharacteristics of listening situations wherein the primary talkerwas in front of them using repeated in situ surveys. The researchparticipants reported the location of noise and the availability ofvisual cues in each situation. However, because the purpose ofWu and Bentler (2010b) was to examine the effect of visual cueson directional microphone hearing aid benefit, the descriptivestatistics of the listening situation properties were not reported.In a study designed to investigate hearing aid users’ preferencebetween directional and omnidirectional microphones, Waldenet al. (2004) asked adult hearing aid users to report microphonepreference and the properties of major active listening situations using in situ surveys. The questions asked in the surveycategorized the listening environments into 24 unique situations. The categories were arranged according to binary representations of five acoustic factors, including background noise(present/absent), speech location (front/others), and noise location (front/others). The frequency of occurrence of each of the24 unique situations was reported. The most frequently encountered type of listening situations involved the speech from infront of the listener and background noise arising from locations other than the front.Prototype Listening SituationsThe term prototype listening situations, or PLSs, refers toa set of situations that can represent a large proportion of theeveryday listening situations experienced by individuals. Theconcept of a PLS was first introduced by Walden (1997). In particular, Walden et al. (1984) conducted a factor analysis on a selfreport questionnaire and found that there were four dimensionsof hearing aid benefit; one for each unique listening situation.Those unique listening situations included listening to speech inquiet, in background noise, and with reduced (e.g., visual) cues,Copyright 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.

WU ET AL. / EAR & HEARING, VOL. 39, NO. 2, 293–304as well as listening to environmental sounds. Walden (1997)termed these unique listening situations as “PLSs.” Walden andother researchers (Cox et al. 1987b) suggested that hearing aidsshould be evaluated in PLSs so that test results can generalize to the real world. However, the PLSs specified by Walden(1997) do not describe important acoustic characteristics suchas speech level, noise level, and SNR. Further, although previous research has examined the properties of real-world communication situations for adults with hearing loss in terms of SNR(Pearsons et al. 1977; Smeds et al. 2015), availability of visualcues, and speech/noise configuration (Walden et al. 2004), thesedata were individually collected by different studies. Therefore,no empirical data are available for developing a set of PLSs thatcan represent typical speech listening situations and can be usedto create ecologically valid speech-related laboratory testing.Research ObjectivesThe present study had two objectives. The first objectivewas to determine the relationship between speech level, noiselevel, and SNR, as well as the distribution of SNR, in real-worldspeech listening situations for adults with hearing loss, as thedata reported by Pearsons et al. (1977) and Smeds et al. (2015)are not consistent. The second objective was to develop a set ofPLSs that relate to speech listening and describe the (1) SNR,(2) availability of visual cues, and (3) locations of speech andnoise sources in the environments that are frequently encountered by adults with hearing loss. In accordance with the PLSsdescribed by Walden (1997), the PLSs in the present study donot characterize the listener’s intention (e.g., conversation versus focused listening) or the type of listening environment (e.g.,restaurant versus car). However, unlike Walden’s PLSs thatinclude nonspeech sound listening situations, the PLSs in thepresent study only focus on speech listening situations.The present study was part of a larger project comparing theeffect of noise reduction features in premium-level and basiclevel hearing aids. The participants were older Iowa and Illinoisresidents with symmetric mild to moderate hearing loss. Theparticipants were fit bilaterally with experimental hearing aids.During the field trial of the larger study, the participants carrieddigital audio recorders to continuously record environmentalsounds, and they repeatedly completed in situ surveys on smartphones to report the characteristics of the listening situations.SNRs were derived using the audio recordings. SNRs and survey data were then used to develop the PLSs.295previous hearing aid experience. A participant was consideredan experienced user if he/she had at least 1 year of prior hearingaid experience immediately preceding the study. While 20 participants completed the study, two participants withdrew fromthe study because of scheduling conflicts (n 1) or unwillingness to record other people’s voices (n 1).Hearing Aids and FittingIn the larger study, participants were fit with two commercially available behind-the-ear hearing aids. One model was amore expensive, premium-level device and the other was a lessexpensive, basic-level device. The hearing aids were coupledto the participants’ ears bilaterally using slim tubes and custom canal earmolds with clinically appropriate vent sizes. Thedevices were programmed based on the second version of theNational Acoustic Laboratory nonlinear prescriptive formula(NAL-NL2; Keidser et al. 2011) and were fine-tuned accordingto the comments and preferences of the participants. The noisereduction features, which included directional-microphone andsingle-microphone noise reduction algorithms, were manipulated (on versus off) to create different test conditions. All otherfeatures (e.g., wide dynamic range compression, adaptive feedback suppression, and low-level expansion) remained active atdefault settings. The volume control was disabled.Audio RecorderTo derive the SNR, the language environment analysis(LENA) digital language processor (DLP) system was used torecord environmental sounds. The LENA system is designedfor assessing the language-learning environments of children(e.g., VanDam et al. 2012), and the LENA DLP is a miniature,light-weight, compact, and easy-to-use digital audio recorder.The microphone is integrated into the case of the DLP. Duringthe field trial of the study, the DLP was placed in a carryingpouch that had an opening for the microphone port. The pouchMATERIALS AND METHODSParticipantsTwenty participants (8 males and 12 females) were recruitedfrom the community. Their ages ranged from 65 to 80 years witha mean of 71.1 years. The participants were eligible for inclusion in the larger study if their hearing loss met the followingcriteria: (1) postlingual, bilateral, sensorineural type of hearingloss (air-bone gap 10 dB); (2) pure-tone average across 0.5,1, 2, and 4 kHz between 25 and 60 dB HL (ANSI 2010); and(3) hearing symmetry within 20 dB for all test frequencies. Thelarger study focused on mild to moderate hearing loss becauseof its high prevalence (Lin et al. 2011). The mean pure-tonethresholds are shown in Figure 1. All participants were nativeEnglish speakers. Upon entering the study, 15 participants hadFig. 1. Average audiograms for left and right ears of 20 study participants.Error bars 1 SD.Copyright 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.

296WU ET AL. / EAR & HEARING, VOL. 39, NO. 2, 293–304was worn around the participants’ necks so that the microphonelaid at chest height, faced outward, and was not obscured byclothing. The LENA DLP was selected because of its superiorportability and usability. Audio recorders that are easy to carryand use were required because audio data was collected over alonger period (weeks) to better characterize real-world listeningsituations that differ considerably between and within individuals. Note that although the LENA system includes software thatcan automatically label recording segments off-line accordingto different auditory categories, the results generated by theLENA software were not used in the present study.The electroacoustic characteristics of three LENA DLPs,which consisted of 10% of the DLPs used in the study, wereexamined in a sound-treated booth. A white noise and a pinknoise were used as stimuli, and both generated similar results.Figure 2A shows the one-third octave-band frequency responseaveraged across the three DLPs relative to the response of aLarson–Davis 2560 ½ inch microphone. Although the responseof the DLPs is higher than the reference microphone by 6.3 dBat 6 kHz, the response is fairly flat ( 2 dB) between 100 and3000 Hz. Figure 2B shows the broadband sound level measuredusing the DLPs (averaged across two stimuli and three DLPs)as a function of the actual level. It is evident from the figure thatthe DLP has an output limiting algorithm for sounds higher thanapproximately 80 dBA and a low-level expansion algorithm forsounds lower than approximately 50 dBA. The expansion ratioFig. 2. Frequency response (A) and the relationship between the measuredand actual level (B) of the digital audio recorder.is approximately 0.4:1. The effect of the expansion was takeninto account when analyzing data (see the data preparation section below). The DLP is fairly linear for sounds between 50 and80 dBA. Because of the noise floor of the device, the lowestlevel of sound that the DLP can measure is 40 dBA.In Situ SurveyThe EMA methodology was used to collect the informationregarding availability of visual cues and the speech/noise location of real-world listening situations. EMA employs recurringassessments or surveys to collect information about participants’ recent experiences during or right after they occur in thereal world (Shiffman et al. 2008). In the present study, the EMAwas implemented using Samsung Galaxy S3 smartphones. Specifically, smartphone application software (i.e., app) was developed to deliver electronic surveys (Hasan et al. 2013). Duringthe field trial, the participants carried smartphones with them intheir daily lives. The phone software prompted the participantsto complete surveys at randomized intervals approximatelyevery two hours within a participant’s specified time window(e.g., between 8 A.M. and 9 P.M.). The 2-hr interprompt interval was selected because it seemed to be a reasonable balancebetween participant burden, compliance, and the amount ofdata that would be collected (Stone et al. 2003). The participants were also encouraged to initiate a survey whenever theyhad a listening experience they wanted to describe. Participantswere instructed to answer survey questions based on their experiences during the past five minutes. This short time windowwas selected to minimize recall bias. The survey assessed thetype of listening activity (“What were you listening to?”) andprovided seven options for the participants to select (conversations 3 people/conversations 4 people/live speech listening/media speech listening/phone/nonspeech signals listening/not actively listening). The participants were instructed to onlyselect one activity in a given survey. If involved in more thanone activity (e.g., talking to friend while watching TV), the participants were asked to select the activity that happened mostof the time during the previous five minutes. Selection of onlythe primary activity when completing a survey stemmed froma goal of the larger study to develop algorithms that can useaudio recordings to automatically predict listening activitiesreported by participants. The survey also assessed the type oflistening environment (“Where were you?”, home 10 people/indoors other than home 10 people/indoors crowd of people 10 people/outdoors/traffic). The listening activity and environment questions were adapted from Wu and Bentler (2012).Whenever applicable, the survey questions then assessed thelocation of speech signals (“Where was the talker most of thetime?”, front/side/back), availability of visual cues (“Could yousee the talker’s face?”, almost always/sometimes/no), noisinesslevel (“How noisy was it?”, quiet/somewhat noisy/noisy/verynoisy), and location of noise (“Where was the noise most of thetime?”, all around/front/side/back). In the survey, the participants also answered a question regarding hearing aid use duringthat listening event (yes/no). For all questions, the participantstapped a button on the smartphone screen to indicate theirresponses. The questions were presented adaptively such thatcertain answers determined whether follow-up questions wouldbe elicited. For example, if a participant answered “quiet” inthe noisiness question, the noise location question would notCopyright 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.

WU ET AL. / EAR & HEARING, VOL. 39, NO. 2, 293–304be presented and “N/A” (i.e., not applicable) would be assignedas the answer. After the participants completed a survey, theanswers to the questions and the time information were saved inthe smartphone. The survey was designed for the larger study,but only the questions that are relevant to the present study arereported in this article. See Hasan et al. (2014) for the completeset of survey questions.ProceduresThe study was approved by the Institutional Review Boardat the University of Iowa. After agreeing to participate and signing the consent form, the participants’ hearing thresholds weremeasured using pure-tone audiometry. If the participant met allof the inclusion criteria, training regarding the use of the LENADLP was provided. Attention was focused on instructing theparticipants on how to wear the DLP, especially regarding theorientation of the microphone and the pouch (e.g., to alwayskeep the microphone facing outward and not under clothing).The participants were asked to wear the DLP during their specified time window in which the smartphone delivered surveys.The storage capacity of a DLP is 16 hours, and so the participants were instructed to wear a new DLP each day. Each of theDLPs were labeled with the day of the week corresponding tothe day that it was to be worn. If they encountered a confidentialsituation, the participants were allowed to take off the DLP. Theparticipants were instructed to log the time(s) when the DLPwas not worn so these data would not be analyzed.Demonstrations of how to work and care for the smartphone,as well as taking and initiating surveys, were also provided. Theparticipants were instructed to respond to the auditory/vibrotactile prompts to take surveys whenever it was possible and withinreason (e.g., not while driving). Participants were also encouraged to initiate a survey during or right after they experienceda new listening experience lasting longer than 10 min. Eachparticipant was given a set of take-home written instructionsdetailing how to use and care for the phone, as well as whenand how to take the surveys. Once all of the participants’ questions had been answered and they demonstrated competence inthe ability to perform all of the related tasks, they were senthome with three DLPs and one smartphone and began a threeday practice session. The participants returned to the laboratoryafter the practice session. If a participant misunderstood any ofthe EMA- or DLP-related tasks during the practice session, theywere reinstructed on how to properly use the equipment or takethe surveys.Next, the hearing aids were fit, and the field trial of the largerstudy began. In total, there were four test conditions in the largerstudy (2 hearing aid models 2 feature settings). Each condition lasted five weeks, and the assessment week in which participants carried DLPs and smartphones was in the fifth week.After the fourth condition, the participants randomly repeatedone of the four test conditions to examine the repeatability ofthe EMA data, which was another purpose of the larger study.Six participants of the present study, including one experiencedhearing aid user, also completed an optional unaided condition.Therefore, each participant’s audio recordings and EMA survey data were collected in 5 to 6 weeks across all test conditions of the larger study. Even though the data were collectedin conditions that varied in hearing aid model (premium-levelversus basic-level), feature status (on versus off), and hearing297aid use (unaided versus aided), it was determined a priori thatthe data would be pooled together for analysis, as the effect ofhearing aid on the characteristics of the listening situations wasnot the focus of the present study. More importantly, poolingthe data obtained under rather different hearing aid conditionswould make the findings of the present study more generalizable than had they been obtained under just a single condition.Similarly, although the manner by which a survey was initiatedvaried (app-initiated versus participant-initiated), the surveydata collected using both manners would be pooled. The totalinvolvement of participation in the larger study lasted approximately 6 to 8 months. Monetary compensation was provided tothe participants upon completion of the study.Data PreparationBefore analysis, research assistants manually prepared theaudio recordings made by the LENA DLP and the EMA survey data collected by smartphones. The EMA survey data wereinspected first. Surveys in which the participants indicatedthat they were not listening to speech and surveys of phoneconversations (i.e., conversational partner’s speech could notbe recorded) were eliminated. For the rest of the surveys, theaudio recording 5 minutes before the participant conducting thesurvey was extracted. Research assistants then listened to the5-min recording and judged if it contained too many artifacts(e.g., the DLP was covered by the clothing and recorded rubbingsounds) and was unanalyzable. If the recording was analyzable,the research assistants then tried to identify the participant’svoice and the speech sounds that the participant was listeningto. If they judged that the participant was actively engaged in aconversation or listening to th

2010a, b). Therefore, some speech recognition materials can be presented in an audiovisual modality (e.g., the Connected Speech Test; Cox et al. 1987a). Another example is the loca-tion of speech and noise sources. Because this factor can affect speech understanding (e.g., Ahlstrom et al. 2009) and the ben-