Automated Forced-Choice Tests of Speech Recognition

Abstract

Purpose This project was undertaken to develop automated tests of speech recognition, including speech-recognition threshold (SRT) and word-recognition test, using forced-choice responses and computerized scoring of responses. Specific aims were (1) to develop an automated method for measuring SRT for spondaic words that produces scores that are in close agreement with average pure-tone thresholds and (2) to develop an automated test of word recognition that distinguishes listeners with normal hearing from those with sensorineural hearing loss and which informs the hearing aid evaluation process.

Method An automated SRT protocol was designed to converge on the lowest level at which the listener responds correctly to two out of two spondees presented monaurally. A word-recognition test was conducted with monosyllabic words (female speaker) presented monaurally at a fixed level. For each word, there were three rhyming foils, displayed on a touchscreen with the test word. The listeners touched the word they thought they heard. Participants were young listeners with normal hearing and listeners with sensorineural hearing loss. Words were also presented with nonrhyming foils and in an open-set paradigm. The open-set responses were scored by a graduate student research assistant.

Results The SRT results agreed closely with the pure-tone average (PTA) obtained by automated audiometry. The agreement was similar to results obtained with the conventional SRT scoring method. Word-recognition scores were highest for the closed-set, nonrhyming lists and lowest for open-set responses. For the hearing loss participants, the scores varied widely. There was a moderate correlation between word-recognition scores and pure-tone thresholds which increased as more high frequencies were brought into the PTA. Based on the findings of this study, a clinical protocol was designed that determines if a listener's performance was in the normal range and if the listener benefited from increasing the level of the stimuli.

Conclusion SRTs obtained using the automated procedure are comparable to the results obtained by the conventional clinical method that is in common use. The automated closed-set word-recognition test results show clear differentiation between scores for the normal and hearing loss groups. These procedures provide clinical test results that are not dependent on the availability of an audiologist to perform the tests.

Keywords

forced-choice - speech recognition - speech-recognition threshold

There is a growing interest in the development of methods that reduce the dependence of basic hearing testing on highly trained testers. This interest stems from the goals of increasing access to hearing testing and more efficient use of professional time. Conventional methods used by audiologists for basic testing (pure-tone and speech audiometry) are based on a set of rules that are easily encoded in software. Once the rules are agreed upon, an automated method can implement them in a standardized way, avoiding errors and shortcuts that occur in conventional testing. The quality indicators that are consciously and unconsciously used by expert audiologists to ensure test accuracy can be embedded in the procedure, providing quantitative estimates of accuracy. Automated pure-tone audiometry (AMTAS) has been found to produce results that are equivalent to those obtained by expert audiologists[1] [2] especially when quality indicators are incorporated into the protocol.[3]

Speech recognition is routinely tested in the basic clinical hearing evaluation. Typically, speech materials are presented in quiet or noise in open-set paradigms with the tester scoring the listeners' verbal repetition of the stimuli. Two approaches have been used to automate speech-recognition tests. Automatic speech-recognition software has been used to score the listeners responses. In our experience, automatic speech-recognition produced accurate speech-recognition threshold (SRT) measurements with spondee stimuli but was not adequate to accurately score responses to monosyllabic words spoken by listeners with a wide range of speech production characteristics.[4] Ooster et al (2020)[5] reported that automatic speech recognition accurately scored responses to sentence stimuli, but as their stimuli were uncalibrated, SRTs could not be compared with those obtained with tester scoring. Speech recognizers are sometimes “trained” on the speech characteristics of the speaker, a process that was not employed in our project due to time constraints. Incorporating this training into the method and the improvements in speech recognizers over the past 10 years may result in more accurate scoring.

In another approach, speech stimuli are presented in a closed-set paradigm with the listener selecting the alternative that matches the stimulus. Black (1957)[6] and House et al (1965)[7] developed lists of monosyllabic words with multiple response alternatives from which the listener selected a response from a paper scoresheet to obtain the percent correct. Dewyer et al (2018)[8] compared open-set word-recognition scores obtained in a clinic setting with closed-set scores obtained with a cell phone app that presented words through Apple EarPods. There was a high correlation between the scores obtained by the two methods in spite of the absence of calibration of the iPhone system. As demonstrated by McCullough et al (1992; 1994)[9] [10] that approach is easily automated, replacing the scoresheet with a touchscreen and automating the scoring with software. A speech-recognition test using this approach is described in this report.

Closed-set methods produce word-recognition scores that are substantially different from open-set scores. There are two sources of these differences, one methodological and one cognitive.[11] First, chance performance is determined by the number of response alternatives. In an open-set paradigm, the number of alternatives is the listener's entire lexicon and a 0% score is possible. The theoretical minimum score in a closed-set, four-interval paradigm such as the one used in this study is 25% and the distribution of scores is significantly compressed relative to the open-set condition. Second, the difficulty of a word-recognition task is influenced by lexical competition, the difficulty of discriminating between similar alternatives. In a closed-set task, lexical competition can be increased by using alternatives that are similar to the target word, such as rhyming words. As the number of alternatives and lexical competition increase in a closed-set task, performance approaches that of open-set recognition. Due to large differences in the distributions of scores from the two paradigms, different interpretation guidelines are necessary. In this study, distributions of recognition scores for spondaic and monosyllabic words presented in a four-interval forced-choice paradigm to young listeners with normal hearing and listeners with sensorineural hearing loss are reported. Closed-set scores for monosyllabic words with rhyming and nonrhyming alternatives were measured to vary the lexical competition of the task.

The specific aims of this project were (1) to develop an automated method for measuring SRT for spondaic words that produces scores that are in close agreement with average pure-tone thresholds and (2) to develop an automated test of word recognition that distinguishes listeners with normal hearing from those with sensorineural hearing loss and which informs the hearing aid evaluation process.

Methods

Participants

Thirty adults with normal hearing and 20 adults with sensorineural hearing loss participated in the SRT study. All were native American English speakers. A subset of 10 adults with normal hearing and 16 of the adults with sensorineural hearing loss participated in the word-recognition study. The participants with normal hearing were recruited from the student body of the University of Minnesota, Twin Cities campus with age ranging from 18 to 30 years. The participants with sensorineural hearing loss were recruited from the Audiology Clinic of the Fairview-University Medical Center, Minneapolis, with age ranging from 18 to 79 years. Tympanometry, otoscopy, and bone conduction testing ruled out external and middle ear disorders. The normal and hearing loss groups had equal numbers of male and female participants. As test results from two ears of the same participant are correlated and thus not independent, one ear of each participant was tested. The right or left ear was randomly selected for the normal group. For the hearing loss group, the better ear was selected to minimize the need for contralateral masking. The protocol was approved by the University of Minnesota Institutional Review Board and all participants signed consent forms.

Pure-Tone Audiometry

All participants were tested with AMTAS.[2] The participants in the normal group had audiometric thresholds that were less than or equal to 20 dB HL at octave frequencies from 250 to 8,000 Hz. Participants with hearing loss had a wide range of audiometric severities and patterns and four-frequency (500, 1,000, 2,000, and 4,000 Hz) pure-tone averages (PTAs) that ranged from 21 to 59 dB HL (mean = 40 dB HL). To facilitate a description of the widely varying audiometric patterns of the hearing loss group, audiograms were classified with automated method for classification of audiograms (AMCLASS), a method for classifying audiograms by severity, configuration, and site of lesion.[12] AMCLASS categories were validated against responses of a panel of expert judges, who categorized the severity, configuration, and site of lesion of a set of 200 audiograms with varying patterns. AMCLASS rules were derived that maximize the agreement between AMCLASS categories and those of the expert judges. A large clinical database of 27,554 audiograms was analyzed[13] and average audiograms were calculated for each configuration category. The average audiograms are based on the number of cases in the database that fell into the various categories and ranged from 166 cases for the moderate trough configuration to 4,006 cases for the mild–severe sloping configuration. [Fig. 1] shows average audiograms for the configurations of audiograms of the hearing loss participants in this study.[13] The audiograms fell into seven severity categories, shown in the legend in the figure with the number of participants in each category.

Fig. 1 Average audiograms for participants in each of seven automated method for classification of audiograms categories. The number of participants in each group is given in the artwork.

Three methods for calculating PTA from the audiograms were used: three-frequency average (500, 1,000, and 2,000 Hz); two-frequency average (500 and 1,000 Hz); and Fletcher rule (average of the lowest two thresholds among 500, 1,000, and 2,000 Hz).[14] [15]

Word-Recognition List Construction

Spondee List

Thirty-one recorded spondaic words from the Auditec of St. Louis compact disc were used to measure SRT. From the original list of 36 words,[16] 5 were omitted based on the subjective judgment of the investigators that they are not in common use (drawbridge, duckpond, hothouse, inkwell, and whitewash). The order of the words was randomized during the SRT protocol, which is described later.

Monosyllabic Word Lists

The 200, Northwestern University Auditory Test No. 6 (NU-6), consonant-nucleus-consonant monosyllabic words[17] spoken by a female (VA-1 speaker)[18] served as the pool of words from which 100 words were selected and organized into four equivalent 25-word lists. Two criteria were used to select the words. First, each word was required to have at least three other words in common usage (based on the subjective judgment of the investigators) that rhymed. This was accomplished by substituting the first letter of each word with the remaining consonants in the alphabet that produced a sizable list of alternative words that was supplemented with a few additional words from online sources. Each of the 100 test words was used three times as a nonrhyming alternative word. Second, the words that were easiest and most difficult on the recognition task by older listeners with sensorineural hearing loss were excluded. The database on which this decision was made was compiled at the Bay Pines VA Medical Center (Table SM11 in the supplemental materials of Wilson & McArdle, 2015)[19] with the VA-1 female speaker recorded version of NU-6 presented monaurally at a comfortable listening level (30–40 dB above the SRT) to 953 patients (mean age = 69.9 years; mean PTA = 34.9 dB HL) during their audiologic evaluations. This produced several test scores for the four, 50-word lists that ranged from 187 to 302, which provided a sufficient sampling of each of the 200 NU-6 words within this older patient population. A target performance range of 60 to 80% correct for each word was selected with the final range being 59.1 to 82.4%. Listings of the four, 25-word lists along with percent correct for each word from the Bay Pines database, the three non-rhyming alternatives are provided in the Appendix ([Tables A1]–[A4]). All four lists, 100 words in total, were presented at each level in the protocol described later.

Speech-Recognition Thresholds

Digitized spondee words were delivered by a prototype audiometer[20] to circumaural earphones (Sennheiser HDA200). The audiometer was calibrated to ANSI S3.6 (2018)[21] specifications for pure-tone and speech stimuli. After each spondee presentation, the listener viewed a touch screen that displayed four textual spondee words in a 2 × 2 matrix with the position of the presented word randomly placed in one of the four matrix positions and the listeners touched the word they thought they heard.

The following bracketing procedure was used to determine the SRT. Two spondees were presented at a starting level of 20 dB re: three-frequency PTA (500, 1,000, and 2,000 Hz). If an incorrect response occurred, the level was increased in 10-dB steps (to a maximum of 80 dB HL) with two words presented at each level until both responses were correct. The level was decreased in 10-dB steps with two words presented at each level until one or two responses at a given level were incorrect. The level was increased in 5-dB steps with two words presented at each level until both responses were correct at a given level. The level was decreased in 5-dB steps with two words presented at each level until at least one incorrect response occurred at a given level. The level was increased in 5-dB steps with two words presented at each level until there was an incorrect response. The SRT was defined as the level in an ascending series at which both responses were correct.

Word-Recognition Test

Word-recognition scores were obtained for three conditions in the following order:

1. Closed-set rhyming—The listener chose a response from the touchscreen which displayed the correct response (in random position) and three alternatives that were words that rhyme with the test word, for example (correct response in caps), SHIRT hurt dirt skirt.

2. Closed-set nonrhyming—The listener chose a response from the touchscreen which displayed the correct response (in random position) and three alternatives that were words that do not rhyme with the test word, for example (correct response in caps), BURN week young hash.

3. Open set—The listener repeated each word and the tester scored the response as correct or incorrect.

Words were presented at the following levels. Levels were selected based on pilot data to capture most of the performance–intensity function, with emphasis on the top half of the function. All levels are referenced to the PTA calculated by the Fletcher rule. All 100 test words were presented at each level. Levels were presented in order from low to high to minimize learning effects.

Normal Group

• Open set—11, 17, 23, 29, 35, 41 dB re: PTA.

• Closed set—6, 10, 14, 18, 22, 26 dB re: PTA.

Hearing Loss Group

• Open set—22, 28, 34, 40, 46 dB re: PTA.

• Closed set—14, 18, 22, 26, 30 dB re: PTA.

Results and Discussion

Speech-Recognition Threshold

Mean PTA, SRT, and SRT–PTA differences for the two subject groups are shown in [Table 1]. For the normal group, mean SRT was lower than PTA with the Fletcher rule producing the smallest differences. The means for the three PTA measures were not statistically different for the normal group (F[2, 87] = 0.55, p = 0.58). The three mean PTA values for the hearing loss group were statistically different (F[2, 57] = 11.8, p = 0.0001). Post hoc testing indicated that the differences between means were significant for two-frequency and three-frequency methods and for three-frequency and Fletcher rule methods and not for two-frequency and Fletcher rule methods at 0.05 confidence level. SRT was lower than PTA for the three-frequency average but not for the two-frequency average and Fletcher rule method.

Table 1 Average 3 Freq PTA, 2 Freq PTA, SRT, and differences between SRT and PTA for three PTA calculation methods

	PTA			SRT	SRT − PTA
	3 Freq	2 Freq	Rule		3 Freq	2 Freq	Rule
Normal
Mean	2.3	1.9	1.1	−0.8	−3.1	−2.7	−1.9
SD	4.4	4.5	4.7	4.4	3.8	4.0	3.6
Hearing loss
Mean	31.0	24.6	24.1	24.3	−6.7	−0.3	0.2
SD	12.0	12.7	13.2	13.6	5.2	3.6	4.0

Abbreviations: Freq, frequency; PTA, pure-tone average; SD, standard deviation; SRT, speech-recognition threshold.

Note: 3 Freq = average threshold at 500, 1,000, and 2,000 Hz; 2 Freq = average threshold at 500 and 1,000 Hz; Rule = average of lowest two thresholds among 500, 1,000, and 2,000 Hz (Fletcher rule).

Distributions of differences between SRT and PTA are shown in [Fig. 2] for the normal group (left panel) and for the hearing loss group (right panel). The figures show the proportion of differences between SRT and the various PTAs for each 5-dB interval. For the normal group, the proportion of differences that were within 5 dB of equality was 87% for two-frequency average, 90% for three-frequency average, and 93% for the Fletcher rule average. For the hearing loss group, the proportion of differences that were within 5 dB of equality was 100% for two-frequency average, 65% for three-frequency average, and 95% for the Fletcher rule method.

Fig. 2 SRT–PTA differences for normal participants (left panel) and hearing loss participants (right panel) for two-frequency PTA (500 and 1,000 Hz) and three-frequency PTA (500, 1,000, and 2,000 Hz) and the Fletcher rule. PTA, pure-tone average; SRT, speech-recognition threshold.

Several reports have derived formulas by regression analysis that predict the SRT by weighted averages of pure-tone thresholds at various combinations of frequencies.[14] [15] [22] [23] [24]. Fletcher[14] [15] suggested that the average of the lowest two thresholds among 500, 1,000, and 2,000 Hz provides good agreement with the SRT. Siegenthaler and Strand (1964)[25] and Carhart (1971)[26] examined various formulas and concluded that a two-frequency average (500 and 1,000 Hz) provided good agreement with SRT. More complicated formulas with weighted averages at more frequencies provided only slight improvement. Carhart (1971)[26] recommended the two-frequency average with a correction factor of −2 dB. In the current study, the Fletcher rule of using the two lowest thresholds did not provide significantly better agreement with the SRT than the two-frequency average. The results indicate that the automated forced-choice procedure produces PTA–SRT differences that agree closely with results of studies that obtained SRT with the commonly used open-set method.

Word-Recognition Scores

Mean word-recognition scores for the normal group are shown in [Fig. 3] with comparison data from Wilson and Antablin (1980)[27] and Wilson et al (1990).[28] Closed-set scores with nonrhyming alternatives are higher than those with rhyming alternatives due to the greater lexical competition of the rhyming alternatives.[11] This is consistent with the observation by Wilson and Antablin (1982)[29] that closed-set word-recognition performance decreases as the similarity of the response alternatives increases. Analysis of variance for the closed-set scores indicated that the effects of level (F[1, 5] = 29.9, p = 0.001) and condition (rhyming vs. nonrhyming) (F[1, 10) = 13.6, p = 0.01) are statistically significant. The closed-set rhyming scores are slightly lower than those reported by Wilson and Antablin (1980)[27] for 24 normal hearing participants using similar, but not identical, stimuli. Open-set scores were compared with the results of the Wilson et al (1990)[28] study that tested 24 participants with normal hearing with similar stimuli. Their results were referenced to dB SPL. The comparison in [Fig. 3] is based on an assumption of an average PTA of 5 dB HL for the listeners in the Wilson et al's (1990)[28] study. The slopes of the two functions at 50% point are identical (4.3%/dB).

Fig. 3 Mean word-recognition scores for the normal group for open-set and closed-set paradigms. The closed-set paradigms utilized rhyming and nonrhyming alternatives. Vertical lines are 1 SD for the data in the current study. Data from Wilson et al (1990)[28] and Wilson and Antablin (1980)[27] are shown for comparison. SD, standard deviation.

Individual word-recognition scores for the rhyming closed-set condition for the hearing loss group are shown in [Fig. 4]. The bold lines represent the normal range defined by the means for the normal group ± 2 standard deviation (SD) with a maximum of 100%. Scores for the hearing loss group varied widely. Some were within the normal range at all levels (e.g., S5 and S14); some were outside the normal range at all levels (e.g., S19 and S20); some were within the normal range at low levels and fell outside the normal range at higher levels (e.g., S16 and S17).

Fig. 4 Closed-set word-recognition scores (rhyming alternatives) for 16 participants with sensorineural hearing loss. The bold lines are the normal range (mean score of normal group + 2 SD with a maximum of 100%). PTA, pure-tone average; SD, standard deviation.

Individual open-set word-recognition scores for hearing loss listeners are shown in [Fig. 5]. The bold lines represent the normal range defined by the means for the normal group ± 2 SD. Similar to the closed-set data, about half of the hearing loss group fell within the normal range.

Fig. 5 Open-set word-recognition scores for 16 participants with sensorineural hearing loss. The bold lines are the normal range (mean score of normal group + 2 SD with a maximum of 100%). PTA, pure-tone average; SD, standard deviation.

[Fig. 6] shows the relationship between closed-set and open-set scores for the hearing loss participants averaged across all presentation levels. Although the range of closed-set scores is substantially compressed relative to the open-set scores, the high correlation (0.87) suggests that the two methods measure similar auditory processes.

Fig. 6 Closed-set (rhyming alternatives) versus open-set word-recognition scores for hearing loss listeners. Each datum point is the closed-set score averaged for all presentation levels plotted against the mean open-set score for an individual subject.

The possibility that participants who fell within an AMCLASS group would show less variability than the group as a whole was explored by examining scores within groups. [Fig. 7] shows individual scores for participants with normal-to-moderate sloping hearing loss. It is clear that the audiometric pattern does not account for the wide variability of the speech-recognition results. Similarly, age does not account for the variability. The age range of the participants with normal-to-moderate sloping hearing loss was 46 to 72 years. The correlation between age and average open-set score across all presentation levels score was −0.07.

Fig. 7 Open-set word-recognition scores for four participants with normal-to-moderate sloping sensorineural hearing loss. The bold lines are the normal range (mean score of normal group + 2 SD with a maximum of 100%). Audiograms for the four participants are shown in the bottom panel. PTA, pure-tone average; SD, standard deviation.

Substantial differences in open-set and closed-set scores were observed for recognition of monosyllabic words for normal ([Fig. 3]) and hearing loss ([Figs. 5] and [7]) listeners. Closed-set testing compressed the range of scores relative to open-set testing. In a closed-set four-interval forced-choice test, average minimum correct rate (i.e., when the stimulus is not audible) is expected to be 25%, whereas for open-set testing, a minimum score of 0% is possible. This accounts for the large differences in the distributions, shown in [Figs. 5] and [6].

The wide range of variability for the open-set scores of the hearing loss group shown in [Fig. 5] is striking. The scores in [Fig. 5] are similar to open-set results with the same speech materials reported by Wilson (2011)[30] and to scores obtained with word lists that were similarly constructed reported by Margolis and Millin (1971).[31] In general, the scores are lower than those obtained with recordings of a male speaker. To further compare word-recognition scores for the hearing loss participants obtained with male and female speakers, scores from clinical evaluations (male speaker) were compared with scores obtained in this study (female speaker). The clinic tests were performed with the Q/MASS recordings of NU-6 words by a male speaker.[32] For this comparison, participants were selected when there were clinic tests performed at levels within 3 dB of a level used in this study. [Fig. 8] shows the open-set scores for those participants along with the clinic scores. The scores obtained with recordings made by a female speaker are substantially lower and more variable than those obtained with the male-speaker recordings.

Fig. 8 Open-set word-recognition scores for the hearing loss participants obtained with two sets of recordings of NU-6 monosyllabic words. The clinic scores were obtained in a clinical hearing evaluation using the Q/MASS recordings of male speech. The AMTEST scores were those obtained in this study with VA recordings of female speech. Error bars are +1 standard deviation.

The results in [Fig. 7] for hearing loss listeners with normal-to-moderate sloping hearing losses suggest that neither audibility nor age accounts for the wide range of variability. To further explore the relationship between word-recognition scores and audibility, correlation coefficients were calculated between average open-set word-recognition scores at all presentation levels and PTAs and between word-recognition scores and thresholds at individual frequencies. These are shown in [Fig. 9]. The correlation between average open-set word-recognition score and PTAs increased substantially as more high-frequency thresholds were included in the average. The correlations with pure-tone thresholds progressed from −0.03 at 500 Hz to −0.24 at 1,000 Hz to −0.73 at 2,000 Hz and then decreased slightly to −0.67 at 4,000 Hz. The decrease at 4,000 Hz is consistent with the lower importance of that frequency relative to 2,000 Hz for recognition of monosyllabic words.[33] Taken together, the data in [Figs. 8] and [9] show evidence of the contributions of both the distortion component and the attenuation component of hearing loss discussed by Plomp (1978),[34] referred to by Carhart (1951)[35] as the acuity component and clarity component.

Fig. 9 Pearson's product-moment correlation coefficients between average open-set word-recognition score and pure-tone averages (2 Fr = 500, 1,000 Hz; 3 Fr = 500, 1,000, 2,000 Hz; 4 Fr = 500, 1,000, 2,000, 4,000 Hz) and between average open-set word-recognition score and thresholds at 1,000, 2,000, and 4,000 Hz.

The closed-set scores in [Fig. 4] indicate that most of the scores of hearing loss listeners are in the normal range at low levels and most are outside the normal range at high levels. The results indicate that, as a group, the hearing loss listeners have significant deficits in recognition of monosyllabic words presented in quiet. The results challenge the widely held notion that listeners with sensorineural hearing loss have little difficulty in quiet but speech recognition in noise brings out the deficits that produce speech communication difficulties.

A Clinical Protocol

A clinical protocol should be based on the information that is desired from a test and rules for how the information is to be interpreted. A clinical protocol has been designed for an automated closed-set speech-recognition test that addresses two questions.

1. Is the listener's speech-recognition performance in the normal range?

2. Does an increase in speech level result in improved performance?

The protocol differs from the usual practice of obtaining a score at a high level to determine maximal performance (“PB-max”).

To address the first question, ideally, scores would be obtained at a wide range of presentation levels. This is not feasible in a clinical regimen. The proposed protocol examines the score at 22 dB re: two-frequency PTA. At this level, hearing loss listeners in our sample whose scores were in the normal range had scores in the normal range at all levels ([Fig. 4]). Hearing loss listeners whose scores were outside the normal range had scores outside the normal range at all higher levels. For our sample, the test performed at 22 dB re: two-frequency PTA effectively separates listeners with normal hearing from listeners with sensorineural hearing loss.

The second question may provide information that is useful for evaluating hearing aid candidacy and for predicting hearing aid success. The protocol calls for testing at a second level (30 dB re: PTA) if the first score is outside the normal range. If the score does not increase at the second level, hearing aid success may be limited and a high gain setting may be inadvisable. In those cases, other rehabilitative methods, such as auditory training and communication strategies, should be considered.

[Fig. 10] shows results of this clinical protocol for all the hearing loss participants. The normal range, defined as the mean ± 2 SD (with a maximum of 100%), is shown by the dashed lines. These plots illustrate the position of the score relative to normal and the effect of increasing the stimulus level. In some cases, the procedure may not capture the maximum score that would be obtained at a higher level. A test could be performed at a third level to obtain the maximum score. Further evaluation of this clinical protocol is needed to evaluate its diagnostic utility and value in the hearing aid selection process.

Fig. 10 Closed-set (rhyming alternatives) word-recognition scores (solid lines) for monosyllabic words presented at 22 and 30 dB (re: PTA) for 16 participants with sensorineural hearing loss. The dashed lines show the normal range. The age (years) is shown in the lower left corner of each panel.

Limitations

The materials reported here were validated on listeners who are native American English speakers. The forced-choice method requires a level of literacy that permits identification of written words. Words were presented in the absence of background noise and the tests have not been validated for speech in noise conditions.

Conclusion

The automated forced-choice SRT procedure described in this report produces threshold values that are in close agreement with PTAs. The Fletcher rule PTA (average of best two thresholds among 500, 1,000, and 2,000 Hz) provided the best agreement for hearing loss participants.

The forced-choice word-recognition test resulted in distinct separation of scores from the normal and hearing loss groups. A clinical protocol was designed that provides information about deviations from normal scores and the possible benefits of increasing the presentation level.

Conflict of Interest

None declared.

Acknowledgment

We are grateful to Dr. Pamela Schreiner for assistance with the statistical analysis in this project.

Note

This work was presented at the annual meeting of the American Auditory Society, March 4, 2016.

Disclaimer

Any mention of a product, service, or procedure in the Journal of the American Academy of Audiology does not constitute an endorsement of the product, service, or procedure by the American Academy of Audiology.

• References

• 1 Mahomed F, Swanepoel W, Eikelboom RH, Soer M. Validity of automated threshold audiometry: a systematic review and meta-analysis. Ear Hear 2013; 34 (06) 745-752
• 2 Margolis RH, Glasberg BR, Creeke S, Moore BCJ. AMTAS: automated method for testing auditory sensitivity: validation studies. Int J Audiol 2010; 49 (03) 185-194
• 3 Margolis RH, Saly GL, Le C, Laurence J. Qualind: a method for assessing the accuracy of automated tests. J Am Acad Audiol 2007; 18 (01) 78-89
• 4 Margolis RH. Automated speech recognition test. Presented to the American Auditory Society, Scottsdale, AZ; March 4, 2010
• 5 Ooster J, Krueger M, Bach J-H, Wagener KC, Kollmeier B, Meyer BT. Speech audiometry at home: automated listening tests via smart speakers with normal-hearing and hearing-impaired listeners. Trends Hear 2020; 24 (04) 1-13
• 6 Black JW. Multiple-choice intelligibility tests. J Speech Hear Disord 1957; 22 (02) 213-235
• 7 House AS, Williams CE, Heker MH, Kryter KD. Articulation-testing methods: consonantal differentiation with a closed-response set. J Acoust Soc Am 1965; 37 (01) 158-166
• 8 Dewyer NA, Jiradejvong P, Henderson Sabes J, Limb CJ. Automated smartphone audiometry: Validation of a word recognition test app. Laryngoscope 2018; 128 (03) 707-712
• 9 McCullough JA, Cunningham LA, Wilson RH. Auditory-visual word identification test materials: computer application with children. J Am Acad Audiol 1992; 3 (03) 208-214
• 10 McCullough JA, Wilson RH, Birck JD, Anderson LG. A multimedia approach for estimating speech recognition of multilingual clients. Am J Audiol 1994; 3 (01) 19-22
• 11 Clopper CG, Pisoni DB, Tierney AT. Effects of open-set and closed-set task demands on spoken word recognition. J Am Acad Audiol 2006; 17 (05) 331-349
• 12 Margolis RH, Saly GL. Toward a standard description of hearing loss. Int J Audiol 2007; 46 (12) 746-758
• 13 Margolis RH, Saly GL. Distribution of hearing loss characteristics in a clinical population. Ear Hear 2008; 29 (04) 524-532
• 14 Fletcher H. A method of calculating hearing loss for speech from an audiogram. Acta Otolaryngol 1950; a; 38 (Suppl. 90) 26-37
• 15 Fletcher H. A method of calculating hearing loss for speech from an audiogram. J Acoust Soc Am 1950; b; 22 (01) 1-5
• 16 Hirsh IJ, Davis H, Silverman SR, Reynolds EG, Eldert E, Benson RW. Development of materials for speech audiometry. J Speech Hear Disord 1952; 17 (03) 321-337
• 17 Tillman TW, Carhart R. An expanded test for speech discrimination utilizing CNC monosyllabic words (Northwestern University Auditory Test No. 6, Technical Report No. SAM-TR-66–55). Brooks Air Force Base, TX: USAF School of Aerospace Medicine; 1966
• 18 Department of Veterans Affairs. Speech Recognition and Identification Materials. Disc 4.0. Mountain Home, TN: VA Medical Center; 2010
• 19 Wilson RH, McArdle R. The homogeneity with respect to intelligibility of recorded word-recognition materials. J Am Acad Audiol 2015; 26 (04) 331-345
• 20 Margolis RH. A self-calibrating audiometer. Presented to the American Auditory Society, Scottsdale, AZ; March 8, 2012
• 21 American National Standards Institute. ANSI S3.6–2018. Specification for Audiometers. New York, NY: American National Standards Institute; 2018
• 22 Harris JD, Haines HL, Myers CK. A new formula for using the audiogram to predict speech hearing loss. AMA Arch Otolaryngol 1956; 63 (02) 158-176
• 23 Delk JH, Glorig A, Quiggle RR, Summerfield AB. Predicting hearing loss for speech from pure tone audiograms. Laryngoscope 1957; 67 (01) 1-15
• 24 Wilson RH, Margolis RH. Measurements of auditory thresholds for speech stimuli. In: Konkle DF, Rintelmann WF. eds. Principles of Speech Audiometry. Baltimore, MD: University Park Press; 1983: 79-126
• 25 Siegenthaler BM, Strand R. Audiogram-average methods and SRT scores. J Acoust Soc Am 1964; 36 (03) 589-593
• 26 Carhart R. Observations on relations between thresholds for pure tones and for speech. J Speech Hear Disord 1971; 36 (04) 476-483
• 27 Wilson RH, Antablin JK. A picture identification task as an estimate of the word-recognition performance of nonverbal adults. J Speech Hear Disord 1980; 45 (02) 223-238
• 28 Wilson RH, Zizz CA, Shanks JE, Causey GD. Normative data in quiet, broadband noise, and competing message for Northwestern University Auditory Test No. 6 by a female speaker. J Speech Hear Disord 1990; 55 (04) 771-778
• 29 Wilson RH, Antablin JK. The picture identification task, a reply to Dillon. J Speech Hear Disord 1982; 47 (01) 111-112
• 30 Wilson RH. Clinical experience with the words-in-noise test on 3430 veterans: comparisons with pure-tone thresholds and word recognition in quiet. J Am Acad Audiol 2011; 22 (07) 405-423
• 31 Margolis RH, Millin JP. An item-difficulty based speech discrimination test. J Speech Hear Res 1971; 14 (04) 865-873
• 32 Massachusetts Eye and Ear Infirmary. Q/MASS Speech Audiometry, Vol. 3. Boston MA: Massachusetts Eye and Ear Infirmary; 1991
• 33 Studebaker GA, Sherbecoe RL. Frequency-importance and transfer functions for recorded CID W-22 word lists. J Speech Hear Res 1991; 34 (02) 427-438
• 34 Plomp R. Auditory handicap of hearing impairment and the limited benefit of hearing aids. J Acoust Soc Am 1978; 63 (02) 533-549
• 35 Carhart R. Basic principles of speech audiometry. Acta Otolaryngol 1951; 40 (1-2): 62-71

Publication History

Received: 09 February 2021

Accepted: 01 July 2021

Article published online:
17 February 2022