Development and validation of a digits-in-noise hearing test in Persian

Abstract

OBJECTIVE:

The prevalence of unrecognized and late-diagnosed hearing loss is higher in low- and middle-income than in high-income countries, due in part to lack of access to hearing services. Because hearing screening is important for early identification of hearing loss, development of an accessible, self-screening test that can detect hearing loss reliably and quickly would provide significant benefits, especially for underserved populations. This study aimed to develop and validate a new version of the digits-in-noise (DIN) test for Persian speaking countries.

DESIGN:

Recordings of Persian digits 0–9 were binaurally presented in broadband speech-shaped noise. Using fitted speech intelligibility functions, digits were homogenized to achieve equal perceptual difficulty across stimuli. Evaluation was established by reference to existing English DIN tests.

STUDY SAMPLE:

Thirty Persian speaking young adults with normal hearing thresholds (≤ 20 dB HL, 0.25 – 8 kHz).

RESULTS:

Speech intelligibility functions produced a mean speech reception threshold (SRT) of −7.7 dB, corresponding closely to previously developed DIN tests. There was no significant difference between test-retest SRTs, indicating high reliability of the test. Our findings suggest that language-specific factors need to be considered for cross-language comparison of DIN-SRTs.

CONCLUSIONS:

This study introduces a convenient tool for future hearing screening in Persian speaking countries with limited access to audiology services.

Introduction

Approximately 6.8% of the world’s population, about half a billion people, suffer from disabling hearing loss (Wilson et al, 2017). Since most hearing loss in adults is progressive, gradual, and painless, it is often ignored until several years after its onset (Davis et al, 2007; Mackenzie & Smith, 2009; Trumble & Piterman, 1992). Studies have shown that in high income countries only 20% of adults with hearing loss seek help, and most of them tend to postpone treatment until difficulty with speech understanding interferes with their work and social lives (Davis et al, 2007; Karpa et al, 2010; Trumble et al, 1992). In low and middle income countries (LMICs), limited, and often unavailable audiology services, together with high poverty rates and poor general health, are other significant reasons for the higher prevalence of hearing loss (Swanepoel et al, 2019; Wilson et al, 2017; World Health Organization [WHO], 2019). Persian speaking countries¹, for example, have limited access to audiology services. In Iran, where Persian is the official language, there is around one audiologist for every 40,000 people (Hearing Health & Technology Matters, 2011). That ratio is substantially lower in the other primarily Persian speaking countries, Afghanistan and Tajikistan.

Untreated hearing loss in adults has multiple negative consequences, including poor quality of life, depression, social isolation, and significant costs to society (Chia et al, 2007; Mackenzie et al, 2009; Stika & Hays, 2015). Particularly in LMICs, with a higher proportion of young people suffering from hearing loss (Blum & Boyden, 2018; National Transfer Accounts Bulletin, 2012), it has detrimental effects on the development of communicative and cognitive skills, emotional well-being, and academic success, resulting in limited career opportunities, poverty, and lower life expectancy (Punch et al, 2004; Qi & Mitchell, 2012; Warner-Czyz et al, 2015). Early detection and treatment of hearing loss are important so that steps can be taken to preserve and enhance hearing, and quality of life for affected people (Gratton & Vázquez, 2003; Mehrparvar et al, 2011; Robinson & Sutton, 1979). Hearing screening plays a key role in the early diagnosis of hearing loss (Chia et al, 2007; Davis et al, 2007; Stika et al, 2015).

Pure tone audiometry uses minimal cognitive resources, but is unable to reliably predict the level of difficulty a person will have understanding speech in a challenging environment (Elberling et al, 1989; Smits et al, 2016; Wilson & Weakley, 2004; Zokoll et al, 2012). In contrast, adaptive speech-in-noise tests allow us to reliably and sensitively measure speech recognition difficulties (Miller et al, 1951; Smits et al, 2006; Plomp, 1986; Ramkissoon et al, 2002; De Sousa et al, 2019). There is substantial evidence that digits-in-noise (DIN) screening tests are objective, easily accessible, and can be successfully self-administered via telephone (Smits et al, 2004), internet (Folmer et al, 2017) or smartphone/tablet (Potgieter et al, 2016; Denys et al, 2018) in any moderately quiet setting. DIN tests typically consist of at least 20 presentations of digit triplets (e.g. 7–2-5) embedded in speech-shaped noise. The signal-to-noise ratio (SNR) is varied adaptively to determine the listener’s speech reception threshold (SRT), the SNR at which the intelligibility of triplets is 50% (Smits & Houtgast, 2007; Smits et al, 2013; Vlaming et al, 2014).

DIN tests are useful self-screening tools for many reasons. First, digits are highly familiar stimuli that are easily recognized by a wide range of people, including young children (Koopmans et al, 2018; Moore et al, 2019) and non-native language speakers (Smits et al, 2016). Second, the digits and noise are presented at supra-threshold levels, making the test more relevant to everyday hearing than traditional pure tone threshold tests. There is also no requirement for sound-attenuating booths, since the noise masks lower-level stray sounds, or expensive clinical equipment (De Sousa et al, 2020; Denys et al, 2018; Smits et al, 2013; Vlaming et al, 2014). Third, digit triplet tests have low measurement error and high redundancy compared to other speech materials, such as words and sentences. They may, therefore provide more sensitive SRT estimates (Jansen et al, 2013; Smits et al. 2006). Fourth, the test can be conducted automatically with a simple, closed set of randomized digits, ensuring that performance is maximally influenced by hearing ability rather than other cognitive demands (Smits et al, 2004; Smits et al, 2016; Vlaming et al, 2014). Fifth, estimated SRTs of triplets are stable across a relatively wide range of absolute presentation levels (Jansen et al, 2013), eliminating the need for careful, in situ calibration. Sixth, DIN-SRTs are highly correlated with SRTs of more complex, sentence-based speech-in-noise tests (Bronkhorst & Plomp, 1990; Smits et al, 2013), and with hearing thresholds (De Sousa et al, 2018; Former et al, 2017; Smits et al, 2005; Vlaming et al, 2014). Seventh, the SRT can be measured in a few minutes with a high sensitivity and specificity to detect hearing loss (De Sousa et al, 2020; Koole et al, 2016; Vlaming et al, 2014).

Due to the success of the original Dutch DIN (Smits et al, 2004) and the subsequent increase in identification and treatment of hearing disorders (Smits et al, 2005), the test has been developed in several other languages and dialects including American, Australian, British, Canadian and South-African English, as well as Danish, French, and German.

This study aimed to develop and evaluate a DIN screening test in the Persian language. Introduction of a Persian DIN could offer both personal screening and promote a standardized speech-in-noise test in the clinic. A critical question is whether there are any unique aspects of Persian digits, such as linguistic, spectral, or psychometric properties that need to be considered. Modern Persian shares basic elements of an Indo-European language. However, the development of various fricative (e.g., x, f, θ, ɤ, β, ð) and voiced sibilant (z and ž) sounds in Persian distinguishes it from other Indo-European languages (Emmerick et al, 2016). In contrast to English, for instance, Persian syllables cannot begin with vowels. These language-specific factors question the practical application of developing language-general tests, for example a ‘universal’ speech-in-noise test using only syllables available in all languages (Cameron et al, 2019). In the present study, we compared the Persian DIN thresholds with SRTs of an existing US Midwest English DIN test which was constructed using identical procedures (Motlagh Zadeh et al, 2019). We hypothesized that language-specific factors in the Persian DIN would result in different SRTs compared with the English DIN.

Materials and methods

Participants

Thirty Iranian-Persian speaking participants (Mean = 30 years; SD = 4 years, 16 female) were recruited. They were all graduate students at the University of Cincinnati who had lived in the USA for a period of 3 to 5 years. All participants had pure tone thresholds ≤ 20 dB HL for octave frequencies 0.25 – 8.0 kHz in both ears. Ten participated in a digit homogenization experiment and the remainder participated in an evaluation experiment. All participants gave written informed consent and were reimbursed under approval of the Cincinnati Children’s Hospital Medical Center Institutional Review Board.

Audiological testing

Conventional pure tone audiometry was performed using an Interacoustics Equinox 2.0 audiometer, calibrated to ANSI 3.6 (2010) standards. Participants were tested in a double-walled sound booth (Acoustic Systems, Austin, Texas) meeting criteria of ANSI S3.1 (1999) for audiometric test rooms. Air conduction thresholds were obtained using Sennheiser HDA300 circumaural headphones and the standard Hughson-Westlake method (Carhart & Jerger, 1959).

Digits in noise (DIN) testing

The DIN test was developed based on the methods described by Vlaming et al (2014) and Smits et al (2013).

Speech Stimuli

The first author, a native speaker of Persian (with the Iranian, Farsi dialect² of Persian), recorded the Persian digits. A list of 20 triplets was constructed from 10 digits (0 to 9) where each digit occurred twice at each position. This list was recorded twice at normal vocal effort and speech rate in a sound attenuating booth. Recordings were made through a Neumann TL103 Cardioid microphone, an ART Digital MPA II microphone pre-amplifier, and a TASCAM SS-CDR200 solid state digital recorder at 44.1 kHz sample rate and stored as 16 bit .wav files. Audio editing software (Audacity 2.1.3) was used to cut the individual digits at the zero crossings on visual inspection of the waveform. The most natural sounding digits for each position were selected to create the test triplets. Only triplets with unique, different digits were used. Order of the digits in the triplets was chosen pseudo-randomly to assure that all digits appeared equally often in each position. The occurrence of a standard numerical order of two successive digits was avoided to reduce possible cueing effects (Smits et al, 2013).

Masking Noise

A broadband noise was generated by obtaining the long-term average frequency spectrum across all digits in all positions. The average spectral power in 1/3 octave bands was computed, and a least square function with 25 coefficients was fitted to these average points to generate the frequency response of a broadband filter. The root mean square (RMS) level of the noise was chosen to match the average RMS level of the single digits. Noise started 100 ms before the first digit started. Noise duration was constant (3.25 s) and the inter-digit interval was 175 ms. Mean digit duration was 0.66 sec (SD = 0.09), ranging from 0.51 to 0.83 sec.

Homogenization

To construct a set of equally intelligible digits, the 50% correct SRTs and slopes of the speech intelligibility function of each individual digit need to be calculated (Vlaming et al, 2014). For this purpose, a list of 40 triplets was constructed such that each digit occurred four times at each position (a, b, c). The noise level was fixed at 65 dB SPL, calibrated through the headphones using a Larson Davis 824 sound-level meter and an AEC201 coupler. A customized PsychoPy script (Peirce, 2007) was developed to automatically play the digit triplets in a random order at 10 fixed SNR levels (ranging from −40 dB to +5 dB and in steps of 5 dB) and to store the results. Listeners were seated in a sound booth in front of a computer screen and were asked to enter the responses via a numeric keypad after each triplet was presented. Entering the three digits was a requirement for presentation of the next stimulus. Participants were instructed prior to testing to guess when unable to identify a digit. The test stimuli were played through TDT Psychoacoustic workstations and delivered diotically through Sennheiser (Wedemark, Germany) HD360 Pro headphones. Homogenization testing was completed in one session for each participant, with 400 trials taking about 45–50 minutes. After 200 trials, a short break (3 min) was taken.

Speech intelligibility functions (Equation 1) were fitted to the data to estimate the SRT and the slope of the logistic curve at the SRT per digit per position and for each participant using the following formula:

$$\mathrm{SI}\text{(SNR)}=\frac{1}{1+e^{4s(SRT-SNR)}}$$ (1)

where SI = speech intelligibility (percent correct identification), and s = slope at the SRT. The average SRT across participants for each digit at each position was subtracted from the overall mean SRT across digits for that position to estimate the level corrections required for digit homogenization (Table 1). For this purpose, the level corrections in dB were transformed into a factor to multiply with the digit waveforms to ensure that the psychometric functions of all digits were overlapping at the 50% recognition SNR. The SNR was calculated relative to a reference RMS value calculated as the RMS of all of the pre-homogenization digits concatenated together.

Table 1.

Individual and averaged fitted slopes (%/dB) and SRTs (dB) for the 10 digits at each of the three positions (a, b, c) of the triplet before homogenization

	Position-a			Position-b			Position-c			Average abc
Digits	Slope	SRT	dB-diff	Slope	SRT	dB-diff	Slope	SRT	dB-diff	Slope	SRT
0	6.6	−12.1	0.8	6.8	−10.3	2.6	4.5	−8.7	4.2	6.0	−10.4
1	6.9	−17.5	−4.6	9.2	−13.6	−0.7	7.6	−16	−3.1	7.9	−15.7
2	5.1	−9.9	3.0	7.8	−14.6	−1.7	8.7	−13.1	−0.1	7.2	−12.6
3	8.7	−12.9	0.1	9.5	−14.6	−1.7	7.9	−11.1	1.9	8.7	−12.9
4	6.3	−14.2	−1.3	7.6	−13.9	−0.9	9.0	−15.0	−2.1	7.6	−14.4
5	8.2	−13.5	−0.6	7.8	−14.2	−1.3	8.1	−13.2	−0.3	8.0	−13.7
6	10.1	−12.9	0.0	8.9	−13.2	−0.3	9.8	−13.7	−0.8	9.0	−13.3
7	8.0	−14.7	−1.8	6.9	−10.3	2.6	9.0	−13.1	−0.2	8.2	−12.8
8	9.2	−12.3	0.6	9.4	−11.7	1.2	9.0	−12.5	0.4	9.2	−12.2
9	7.0	−10.6	2.3	7.5	−11.1	1.8	7.7	−12.9	0.0	7.4	−11.6
Average	7.6	−13.1	−0.1	8.2	−12.8	0.2	8.2	−12.9	0.0	7.9	−13.0

dB-diff denotes the difference of each SRT from the (abc) average SRT.

Evaluation

Evaluation is the process of pilot testing the newly homogenized digits and assessing consistency with published studies (Smits et al., 2013; Vlaming et al., 2014). The test was programed to present a random set of 25 homogenized digit triplets diotically through Sennheiser (Wedemark, Germany) HD 25–1 headphones via a Maya 22 USB sound card in a double-walled sound booth (IAC) meeting criteria of ANSI S3.1–1999 for audiometric test rooms. On every trial, three digits were randomly chosen such that the same digit could not appear in more than one position in that trial. There were no constraints on how many times a digit appeared in total or in each position. A 1-up 1-down adaptive procedure with a step size of 2 dB was used to obtain the SRT. All three digits of a triplet had to be identified correctly to count as a correct response. The initial SNR level was –4 dB, about 9 dB above the mean SRT of unhomogenized digits (−13 dB, Table 1) consistent with the initial SNR level used by Smits & Houtgast (2007). The SRT was estimated as the average SNR of the final 19 of 25 total trials. Test-retest reliability and effect of learning were assessed by calculating DIN-SRT twice for each listener. Reference values were determined based on the first test assessment. Evaluations were completed in one 15 minute session.

R software (version 3.4.2) was used for statistical computing and graphics. Two-way Analysis of Variance was used to assess the effects of different digits and digit position on mean slope and SRT. Statistical analysis for between-language (Persian vs English) comparisons used Independent Samples t-Tests. Two-way mixed single intraclass correlation coefficients (ICCs) were used for calculation of test-retest reliability. All tests were two-sided, and a p-value of < 0.05 was set as significance level.

Results

Homogenization

The SRT and slope provided by the fitted speech intelligibility function for each of the digits in a, b, and c positions are shown in Table 1. Position had no significant effect on mean slope and SRT, but SRT differed substantially between digits. Figure 1 shows the group average for correct identification of each digit at each SNR. Average slope steepness of the speech intelligibility functions was 7.9%/dB with a mean SRT = −13 dB.

Figure 1.

Averaged across position and participants speech recognition probabilities for single digits at each SNR before homogenization.

Evaluation

A mean SRT of −7.7 dB (SD = 1.6 dB) was obtained using the adaptive procedure with triplets of homogenized digits, significantly (p <0.0001, t = 3.8, df = 29.1) higher (poorer) than the mean SRT obtained from the US Midwest English DIN test (−9.3 dB, SD = 0.9), developed by Motlagh Zadeh et al (2019) using the same methodology. It was also higher than the mean SRT of the British English DIN test (−10.3 dB, SD =1.1 dB) reported by Vlaming et al (2014). The adaptive procedure data were used for calculation of confusion matrices to assess error patterns between presented digits and responses in Persian (Table 2) and US Midwest English (Table 3). The large differences in probability of correct responses (bold numbers on the diagonal) indicate differences in the intelligibility of the digits. Most of the error probabilities observed in responses were small (<0.05) and were presumably due to listeners guessing when they could not discriminate the digits. However, the larger error probabilities (highlighted in grey in Tables 2 and 3) may have resulted from the digits that are phonetically more similar to each other. For example, Figure 2 illustrates the similar high frequency (including extended high frequency ≥ 8 kHz) energy at the onset of the Persian digits 0-/sefr/ and 3-/se/, and the relatively low amplitude, and low frequency differences at the ends of these digits. Because of the higher acoustic energy in the lower frequencies of the masking noise, such patterns of acoustic similarity and difference are likely to induce perceptual confusions between these digits in the presence of speech-shaped noise. Similarly, increased confusion rates occurred between 7-/hæft/ and 8-/hæʃt/, likely for similar reasons. The confusion pattern of homogenized Persian digits (Table 2) also shows the different pairs between which confusions occurred, and the higher number of confusions of Persian digits compared to the English ones (Table 3). For example, there were five large errors (> 0.1) in Persian digits compared to two such errors in English.

Table 2.

Persian confusion matrix showing the probability of correct responses for each individual digit stimulus. Small probability erroneous responses (0.05 – 0.1) are highlighted in dark grey. Larger probability erroneous responses (≥ 0.1) are highlighted in light grey.

Persian DIN test
Response
		/sefr/	/ yɛk/	/do/	/se/	/tʃǝhar/	/pændɜ/	/ʃeʃ/	/hæft/	/hæʃt/	/noh/
		0	1	2	3	4	5	6	7	8	9
Stimulus	0	0.84	0.00	0.00	0.16	0.00	0.00	0.00	0.00	0.00	0.00
	1	0.00	0.80	0.07	0.01	0.03	0.04	0.00	0.02	0.00	0.02
	2	0.00	0.04	0.94	0.00	0.00	0.00	0.00	0.00	0.00	0.02
	3	0.12	0.00	0.00	0.87	0.00	0.00	0.00	0.00	0.00	0.00
	4	0.00	0.02	0.00	0.00	0.82	0.04	0.00	0.04	0.06	0.01
	5	0.01	0.1	0.02	0.01	0.02	0.73	0.01	0.03	0.01	0.06
	6	0.00	0.04	0.02	0.03	0.04	0.15	0.64	0.03	0.00	0.04
	7	0.00	0.01	0.00	0.02	0.05	0.05	0.00	0.80	0.07	0.00
	8	0.00	0.03	0.04	0.01	0.03	0.19	0.03	0.14	0.52	0.00
	9	0.00	0.01	0.01	0.01	0.00	0.02	0.00	0.00	0.00	0.94

Table 3.

English confusion matrix showing the probability of correct responses for each individual digit stimulus. Small probability erroneous responses (0.05 – 0.1) are highlighted in dark grey. Larger probability erroneous responses (≥ 0.1) are highlighted in light grey.

English DIN test
Response
		/zɪɹo/	/wʌn/	/tu/	/θɹi/	/foʊɹ/	/fʌɪv/	/sɪks/	/sɛvən/	/eɪt/	/naɪn/
		0	1	2	3	4	5	6	7	8	9
Stimulus	0	0.96	0.03	0.00	0.00	0.01	0.00	0.00	0.00	0.00	0.00
	1	0.02	0.84	0.01	0.02	0.02	0.02	0.00	0.01	0.00	0.05
	2	0.01	0.01	0.84	0.1	0.02	0.00	0.00	0.00	0.01	0.00
	3	0.05	0.05	0.02	0.73	0.03	0.01	0.01	0.01	0.09	0.00
	4	0.12	0.09	0.05	0.09	0.55	0.02	0.02	0.01	0.04	0.03
	5	0.03	0.03	0.01	0.03	0.01	0.83	0.00	0.01	0.00	0.04
	6	0.03	0.01	0.00	0.01	0.01	0.00	0.90	0.03	0.01	0.00
	7	0.01	0.00	0.00	0.01	0.03	0.03	0.1	0.81	0.00	0.01
	8	0.03	0.00	0.02	0.03	0.03	0.02	0.02	0.01	0.82	0.02
	9	0.02	0.18	0.00	0.00	0.00	0.04	0.00	0.02	0.00	0.73

Figure 2.

Spectrograms of the Persian digits 0-/sefr/ (a) and 3-/se/ (b).

Test–retest reliability

Figure 3 shows the correlation plot of test-retest SRTs. Mean SRTs slightly improved on retest (Mean = −8.1 dB, SD = 1.9) relative to the first test (Mean = −7.7 dB, SD = 1.6 dB), but this improvement was not statistically significant (t(19) = 1.04, p = 0.3). The 0.4 dB mean SRT improvement on retest suggests a small effect of learning. There was a moderate two-way mixed single intraclass correlation coefficient (ICC = 0.47, p-value = 0.03) between the test-retest SRTs.

Figure 3.

Correlation plot of the Persian digits-in-noise test-retest speech reception thresholds (SRTs). The dashed line shows Test = Retest.

Discussion

Persian DIN development followed similar procedures to those previously used for the construction of the English and Dutch DIN hearing tests (Motlagh Zadeh et al, 2019; Smits et al, 2013; Vlaming et al, 2014). Persian speech intelligibility functions were quite comparable with those of previously developed DIN tests (Jansen et al, 2010; Potgieter et al, 2016; Smits et al, 2016; Vlaming et al, 2014; Zokoll et al, 2012). The mean SRT of the Persian digits, however, was significantly higher, by 1.6 dB, than a US English language version of the test developed following a near-identical protocol (Motlagh Zadeh et al, 2019). This difference may have been related to the small audiometric PTA difference between Persian- and English-speaking participants (1.5 dB) that also favored the English-speaking participants. DIN SRT is highly correlated with audiometric PTA (Jensen et al, 2013; Smits et, 2013; Vlaming et al, 2014). It is therefore likely that a sample with higher PTA would have a higher SRT. The mean SRT of the Persian digits before homogenization (−13 dB) was 5.3 dB better than for homogenized digits (−7.7 dB) at the evaluation stage. This finding is expected, and in agreement with previous DIN studies, because SRTs were calculated for individual digits in the homogenization procedure whereas, during the evaluation phase, responses were only regarded as correct if all three digits were correctly identified.

Precision of threshold estimates may be indicated by the average steepness of the speech intelligibility function slope, which was higher for the Persian (M=7.9 %/dB) than for the English (M=4.4 %/dB) digits. However, both these mean slopes were considerably shallower than those found in other DIN tests (Smits et al, 2013; Vlaming et al, 2014; Zokoll et al, 2012). It seems likely that the broader step size (5 dB) used to derive the functions in this study, compared to that typically used (~ 2 dB) in other studies (Smits et al, 2013; Vlaming et al, 2014), may have contributed to the shallower slopes by lowering the resolution. However, differing slope values may also indicate language-specific variables. For example, in a study by Zokoll et al (2012), the slope of the speech intelligibility functions was steeper for French and Swedish digit triplet tests (27.1 and 24.2 %/dB, respectively) compared to those of Dutch and Greek tests (16.0 and 16.8 %/dB, respectively). Zokoll et al (2012) suggested that languages in which discrimination of digits is more dependent on temporal or consonant cues should have shallower slopes than languages more dependent on spectral cues (or vowel features) because spectrally-matched background noise will have higher masking efficiency for more spectrally-based languages. Persian and English vary from digit-to-digit in the proportion of spectral cues. Comparing spectrograms of Persian digits and US Midwest English digits showed slightly higher spectral cues to Persian digits than English. Koifman et al (2016) also found relatively similar slopes for Semitic languages (Arabic and Hebrew digits), but a shallower slope for Persian digits.

Both the Persian and English tests showed a number of confusions between distinct pairs of digits. In some cases, these confusions seemed predictable based on the phonetic content of the confused digits (e.g. Persian 0 and 3, 7 and 8). However, other digit pairs (Persian 8 and 5, 6 and 5; English 4 and 0) showed an asymmetry of confusion suggesting that, the first digit of each pair was misheard as the second digit. This sort of confusion occurred in both languages to about the same extent, in terms of asymmetry. However, high confusion rates occurred between more pairs of Persian digits than English digits. A possible solution to this might be omitting digits with higher confusion rates from the test, as has been done in some previous DIN tests (Jansen et al, 2010; Vlaming et al, 2014).

Previous studies have shown systematic differences in intelligibility (SRT) between speech in noise tests in different languages (Hochmuth et al, 2015; Kang, 1998; Kollmeier et al, 2015). Kang (1998), for instance, found that the intelligibility of English is better than that of Chinese at the same SNR. Hochmuth et al (2015) also showed language-specific differences between SRTs of native listeners using sentence-in-noise tests in Spanish, German, and Russian. Spanish listeners had higher SRTs than the German or Russian listeners. Hochmuth et al (2015) suggested that the phonological and phonetic properties of Spanish (e.g. less complex syllables, no vowel reduction, and high proportion of sonorant sounds) in contrast to German and Russian may have contributed to this higher vulnerability in noise. However, these studies used ‘open set’ words and complete sentences that have uncertainty and higher linguistic complexity than digits. There are also some other sources of variation, such as different types of noise, that need to be considered across different language SRT comparisons. For example, speech-shaped noise had a less deleterious effect on speech identification than multi-talker babble noise in the study by Hochmuth et al (2015).

Smits et al (2013) argued that understanding digits in noise requires very little linguistic skill (top-down processing) from the listener because digits are among the most frequently spoken words that can be easily learned by children and second language learners. However, there have been very few studies investigating the role of language-specific factors (e.g. different phoneme inventories, and syllable structures) on the intelligibility of the digits. Zokoll et al (2013) developed multilingual (Russian, Turkish, and Spanish) DIN tests and compared them with the other established European DIN tests in British English, Dutch, French, German, Greek, Polish, and Swedish. Their results showed similar speech intelligibility functions for Russian, Turkish, and Spanish DIN tests in contrast to the intelligibility functions of DIN tests in other languages. However, they reported higher mean SRTs for the British English and Russian DIN tests and lower mean SRTs for the Greek and Swedish DIN tests in comparison to the other versions of the tests.

Considering the fact that all these tests provide a highly comparable test format, recording and homogenization procedures, and evaluation standards, differences between the SRTs obtained for the different languages can reasonably be attributed to language- and talker- specific factors (Zokoll et al, 2013). In another study that examined group differences between Dutch and US DIN tests, Smits et al (2016) discussed the role of vowels, consonants, and word length of the digit-triplet stimuli in perceiving digits in noise. They suggested that differences in phonology, speaking rate, articulation and intonation can affect speech perception ability for using DIN tests as a clinical-diagnostic test in different languages.

There is substantial evidence that talker-specific factors (e.g. gender, speaking rate, vowel space, and talker long-term spectrum) significantly affect intelligibility of speech in noise (Bradlow et al, 1996; Hochmuth et al, 2015; Silbert & Motlagh Zadeh, 2018; Smits et al, 2016) and should be considered as an important factor in perceiving speech in noise. Since two different talkers were used for recording Persian and English digits, the contribution of talker-specific factors might also have influenced the observed differences in SRTs. However, Smits et al (2016) showed Dutch-DIN and US-DIN SRTs were highly comparable even though two different talkers of different genders were used for the recording materials. There is a discussion on the possibility of introducing a universal DIN test using an internationally standardized version of an English DIN test, since English is a global language and digits are usually among the first words that are learned in a second language. For example, the World Health Organization recently released the hearWHO app (Swanepoel et al, 2019) that delivers the English version of the DIN test as a self- and community-based hearing screening test. However, since the test is language dependent, it would not have a practical application for widespread global uptake, and especially for many LMICs where knowledge of English is very limited. It is therefore essential that different DIN tests with test-specific normative values are developed in different languages for clinical use. The Persian DIN test described here has the potential to be developed as a website- and/or smartphone-based test to increase access to hearing health care services in Persian speaking countries. The necessary future step is validation of the Persian DIN test in hearing impaired Persian listeners.

Conclusion

A Persian DIN test was developed and evaluated relative to a similarly developed US English DIN test. SRT and slope estimates of the Persian test were obtained for normal hearing, young adult Persian participants. Compared to the US English digits, an increased slope of the digit psychometric functions, a higher SRT, and a different pattern of confusion matrices of the Persian digits suggest that across-language phonological differences should be taken into account when comparing DIN measurements in different languages. As a remotely deliverable screening test, validated in a large, mixed hearing ability sample, a Persian DIN could increase access to hearing services across Persian speaking countries.