Speech Perception in Noise: Do Children Hear Their Peers Differently Than Adults?

Abstract

Background:

Children are constantly exposed to a wide range of speech sources, not only from adults but also from their peers. Yet, the majority of speech perception research has predominantly relied on adult-produced speech as the primary stimulus for pediatric listeners. This oversight is critical, as child and adult speech differ significantly in terms of acoustics and phonetics. These differences raise an important question: How do children process speech produced by both adults and their peers, especially in the presence of background noise? Understanding this dynamic is crucial, as the ability to recognize speech in noisy environments is central to communication and learning. The present study aimed to bridge this gap by investigating how children and adults differ in their recognition of speech-in-noise, particularly when the speech targets are produced by children versus adults. By examining how children process these differences, this work provides invaluable insights into their speech processing mechanisms.

Methods:

The study included 70 typically developing children aged 4 to 17 years (mean age = 9.9) and 27 adults. Speech recognition thresholds (SRTs) were measured for both child- and adult-target talkers in speech-shaped noise. The child target was produced by a 7-year-old girl, while an adult female produced the adult target.

Results:

Analyses revealed that children achieved significantly lower (better) SRTs with child-produced speech compared with adult-produced speech. Interestingly that the effect of the target talker (adult or child) did not differ significantly between listener groups. However, a significant trend emerged showing an interaction between age and target talker. The improvement in SRT with age differed between adult and child talkers, with younger listeners (<10 years) exhibiting a greater advantage for child talkers, which diminished as age increased. A weak correlation between SRTs for child and adult targets suggested that adult-based speech perception tests may not fully represent children’s speech recognition abilities.

Conclusions:

This study underscores the importance of considering child-produced speech in speech perception research, particularly in noisy environments. The developmental differences in speech-in-noise recognition between adult and child talkers suggest that speech perception abilities in children may not follow the same patterns for both talker types. These findings highlight the need for more comprehensive pediatric speech perception studies that include child-target speech stimuli, which could enhance our understanding of speech processing in children and potentially inform interventions and hearing device fittings tailored specifically to the unique needs of pediatric listeners.

INTRODUCTION

Auditory environments of children are rich and diverse, exposing them to a variety of sounds and speakers beyond just adults. While listening to other children makes up a substantial portion of children’s auditory experience, surprisingly little is known about how young listeners process child-produced speech and how it functions as a source of language input. Infants and toddlers in childcare settings spend significant amounts of time hearing their peers, forming the foundation for spoken language development (Bemier & Soderstrom 2016). Similarly, school-age children spend nearly as much time listening to their peers as they do to adults—not just in classrooms but also in hallways, cafeterias, playgrounds, and afterschool programs, where speech produced by their peers may be more prevalent than speech produced by their teachers (i.e., adult talkers) (Lam et al. 2014; Gremmen et al. 2018; Archbell et al. 2020). Furthermore, a significant portion of language learning in children occurs through peer interactions, which provide a natural environment for practicing language skills (Chen et al. 2020; Schmerse 2021; Bushati et al. 2023; Finders et al. 2023). Despite this, nearly all pediatric speech testing relies on speech stimuli produced by adults (Schafer 2010; Holder et al. 2016; Sanchez et al. 2022).

Child speech differs acoustically from adult speech due to physiological and developmental factors (Kent et al. 1994; Potamianos & Narayanan 2007). Because of their smaller and thinner vocal folds, children have a higher fundamental frequency (f₀) compared with adults. On average, children’s voices range between 300 and 500 Hz, while adult female voices range from 180 to 250 Hz and adult male voices range from 80 to 150 Hz (Eguchi & Hirsh 1969; Lee et al. 1999; Baken & Orlikoff 2000). A key difference between adult and child voices can also be observed in vowel formant frequencies (F1, F2, F3). In children, higher formants create a more diffuse vowel space that gradually converges with adult norms as their vocal tract matures (Hillenbrand et al. 1995; Vorperian & Kent 2007). Additionally, child-produced speech contains less low-frequency energy and more high-frequency components, affecting its acoustic profile and potentially impacting intelligibility in noisy environments (Kent et al. 2009). For example, the spectral energy of fricatives produced by infants can extend as high as 16 kHz (Kent & Read 2002).

Beyond spectral characteristics, speech produced by children differs temporally from adult speech. Children generally speak more slowly, producing syllables and words at a lower articulation rate and longer segmental durations than adults (Kent & Forner 1980; Lee et al. 1999; Nissen & Fox 2005; Mahr et al. 2021). In addition, children show greater variability in timing, with phoneme and syllable durations fluctuating more across repetitions (Smith et al. 1996). They also produce longer and more frequent pauses (Redford 2013). Another key difference is their reduced coarticulation, as adjacent sounds are less smoothly connected compared with adult speech, leading to slower transitions between phonemes and syllables (Abakarova et al. 2022). These temporal characteristics contribute to the distinct rhythm and prosody of child speech, which may impact intelligibility, particularly in complex listening environments. Speech produced by children also exhibits greater variability in pitch, intensity, and duration due to immature motor control, making it less stable than adult speech (Smith & Zelaznik 2004). Increased nasality is also common due to incomplete velopharyngeal control (Löfqvist & Gracco 1997; Bunton 2018). Phonetically, children’s speech is characterized by less precise articulation, weaker consonant–vowel contrasts, and developmental phonological processes such as fronting, gliding, and cluster reduction, which gradually resolve with age (Stoel-Gammon & Dunn 1985). In addition, children exhibit greater within-speaker variability in their production of phonemic contrasts, with more dispersed and overlapping phonetic categories compared with adults (Romeo et al. 2013). These differences distinguish child- and adult-produced speech and highlight the need to understand how listeners, particularly children, process speech produced by their peers.

Few studies have investigated how phoneme and word perception in children varies between speech produced by adults and children (Hazan & Markham 2004; Cooper et al. 2018; Bernier & White 2019; Żygis et al. 2023). Hazan and Markham (2004) found that when monosyllabic words were presented in 20-talker babble at +6 dB signal to noise ratio (SNR), speech produced by adult female talkers was slightly more intelligible than that produced by adult males or children (mean age = 13.2 years; 6 males and 6 females), with no significant difference in intelligibility between the adult male and child speech. While younger children (7 to 8 years) exhibited poorer intelligibility than did older children (11 to 12 years) or adults (mean age = 29.9 years), no significant interaction was observed between talker- and listener-groups, suggesting that intelligibility patterns remained consistent across the listener groups. Cooper et al. (2018) showed that toddlers (on average 2.5 years old) were able to recognize words faster and more accurately when speech was produced by a familiar (e.g., mother) or unfamiliar adult female than by another toddler (including their own voice). Similarly, Żygis et al. (2023) observed that children aged ~3 to 8 years were more accurate at identifying bisyllabic words (Polish sibilant contrast/s, ʂ, ɕ/) produced by adults than by their own voice. In contrast, Bernier and White (2019) showed that toddlers (approximately 2 years old) processed familiar words produced by a female 7-year-old child similarly to those produced by an adult female speaker. This sensitivity was largely unaffected by toddlers’ experience with other children. These studies highlight the complexity of perception of child-talkers, with differences depending on listening demands (speech in quiet or noise), the age of the talker, and whether children are listening to their own speech or that of their peers. While some studies found advantages for speech produced by adult females, others suggest that children can process speech produced by adults and children similarly. Taken together, the variability in how children perceive and process speech from different talker ages may depend on the listening context and listening demands.

Little is known about perceptual differences between adult- and child-produced speech for more complex listening conditions. Children often struggle more than adults with understanding speech in noisy environments, and this difficulty can vary depending on the type of background noise (Goldsworthy & Markle 2019; Porto et al. 2023). Moreover, understanding target speech in competing talkers develops more slowly than does understanding speech-in-noise (Leibold & Buss 2019; Leibold et al. 2019). It remains unclear how speech understanding by children in noisy environments is affected by speech produced by adults versus other children. This knowledge gap was the motivation for the present study, which investigated speech understanding in noise by children and adults listening to target speech produced by children and adults. Based on some of the studies cited earlier and the immature acoustic properties of child speech, we hypothesized that children understand speech produced by adults better than speech produced by their peers. Understanding how children perceive speech produced by their peers versus adults holds significant theoretical and clinical implications, potentially contributing to the development of speech-language acquisition models, enhancing pediatric speech assessments, and informing targeted interventions.

MATERIALS AND METHODS

Subjects

The study included 70 typically developing children and 27 adults. The children were aged 4 to 17 years (mean = 9.9), while the adults were aged 18 to 31 years (mean = 21.4 years). All participants had normal audiograms (i.e., hearing thresholds ≤20 dB HL at octave frequencies between 250 and 8000 Hz) and no history of otitis media in the past year. All participants had normal A-type tympanograms, indicating typical middle ear function. In addition, none of the children were receiving any interventions for childhood communication disorders. The adult participants served as a comparison group for speech perception in noise by children. Children were recruited from local school districts in Austin and the surrounding areas. All study procedures were approved by the institutional review board. Experiments were performed in a sound-treated booth.

Audio Recordings and Speech-in-Noise Recognition

Digit-in-noise testing was used to assess speech-in-noise recognition performance. Speech recognition thresholds (SRTs; Plomp & Mimpen 1979) were measured using a closed-set, adaptive (one-up/one-down) procedure, converging on the SNR required for 50% correct recognition (Levitt 1971). SRTs were defined as the SNR needed to achieve 50% correct recognition of a three-digit sequence. Stimuli consisted of the digits “zero” through “nine,” produced by either a 7-year-old female child talker and an adult female talker, both of whom were native speakers of American English. The recordings of all digits were made using an AKG C214 condenser microphone (with a foam cover) through a Focusrite Scarlett 2i2 audio interface, with a 44.1 kHz sampling frequency, and the microphone positioned approximately 6 inches from the speaker’s mouth. For each talker, the speech-shaped steady noise (SSN) masker was spectrally matched to the average spectrum across all 10 digits (i.e., different SSN maskers for the child and adult talkers). Figure 1 displays the average spectrum across all the digits for child and adult talkers. Because the goal of the present study was to determine whether child- or adult-produced speech is intrinsically more robust in noise, different spectrally matched SSN maskers were used for the adult and child talkers. If the long-term average speech spectrum of the masker was not matched to the target talker (i.e., using a single masker for both talkers), one stimulus set may have received an unintended spectral advantage due to less masker-target spectral overlap. As such, different but spectrally matched SSN was used to measure speech understanding with the child and adult talkers.

Fig. 1.

Average spectrum across all 10 digits for adult and child targets. Note that the average spectrum for each talker was used to create the speech-shaped noise masker for each talker.

Table 1 presents duration, fundamental frequency (f₀), and the mean frequency for the first three formants (F1, F2, and F3) for the child and adult female talkers in the present study, as well as published f₀ and formant values from Hillenbrand et al. (1995) and Lee et al. (1999). Across all digits, the mean duration was 522.9 ± 115.6 msec for the child talker and 550.4 ± 101.9 msec for the adult talker, comparable to the duration of Consonant–Nucleus–Consonant (CNC) words used for audiometric speech testing. Across all digits, the mean f₀ across digits was 249 Hz for the child talker and 212 Hz for the adult talker. Directly comparing the vowels common to the present study (ʌ, u, I, ɪ, ɛ), Hillenbrand et al., and Lee et al. [ʌ, u, I, ɪ, ɛ], the mean f₀ for the child talker was 253.8 ± 11.9 Hz for the present study, was 240.4 ± 7.6 Hz for Hillenbrand et al. (1995), and 281.6 ± 9.4 Hz for Lee et al.. Thus, the present child’s mean f₀ falls within the range reported by Hillenbrand et al. and Lee et al.. Compared with the values reported by Hillenbrand et al. for groups of 10 to 12-year-old children and adult female talkers, the present recordings yielded mean frequency ratios (present/previous) of 1.00 ± 0.14 for the child and 0.94 ± 0.20 for the adult talker across f₀, F1, F2, and F3. Similarly, relative to Lee et al., which included data from 7-year-old children (the same age as in the present study), the mean ratios were 0.93 ± 0.13 for the child and 0.98 ± 0.14 for the adult. These results confirm that the formant characteristics of the present stimuli are consistent with established norms for child and adult female speech.

TABLE 1.

Duration (in ms), fundamental and formant frequencies for child and adult talkers compared with published norms

			Zero		One	Two	Three	Four	Five	Six	Seven		Eight	Nine
			/zɪ/	/ro/	/wʌn/	/tu/	/θri/	/fɔr/	/faɪv/	/sɪks/	/sɛ/	/vən/	/eɪt/	/naɪn/
Present study	Child	Duration	683		458	375	528	564	563	650	600		318	490
		f 0	252	224	246	239	254	255	250	261	269	238	263	241
		F1	396	620	755	541	475	705	1034	680	813	673	765	890
		F2	2574	1478	1728	1527	3117	1166	1608	2094	2043	2199	2756	1943
		F3	3456	2636	2723	2278	3687	2522	2244	2977	2558	2675	3425	2321
	Adult	Duration	649		487	470	410	480	619	712	663		482	532
		f 0	226	197	205	220	204	203	208	225	227	203	221	209
		F1	406	477	752	435	383	429	907	516	633	653	496	824
		F2	2300	1169	1463	1823	2597	702	1644	2153	1449	2076	2256	1467
		F3	2765	2566	3034	2771	3105	2332	2564	2955	2539	2994	3047	2394
Hillenbrand et al. (1995)	Child	f 0	241	236	236	249	246	225		241	230
		F1	511	597	749	494	452	803		511	749
		F2	2552	1137	1546	1345	3081	1210		2552	2267
		F3	3403	2987	3145	2988	3702	2982		3403	3310
	Adult	f 0	224	217	218	235	227	210		224	214
		F1	483	555	753	459	437	781		483	731
		F2	2365	1035	1426	1105	2761	1136		2365	2058
		F3	3053	2828	2933	2735	3372	2824		3053	2979
Lee et al. (1999)	Child	f 0			298	281	276			278	275
		F1			661	506	467			608	856
		F2			1566	1840	3026			2642	2428
		F3			3327	3316	3573			3482	3458
	Adult	f 0			246	243	228			235	219
		F1			595	412	360			532	694
		F2			1522	1388	2757			2183	2057
		F3			2887	2804	3291			3064	3005

Fundamental frequency (f₀, in Hz) and formant frequencies (F1–F3, in Hz) for the present digits produced by the child and adult female talkers, as well as mean values from Hillenbrand et al. (1995) and Lee et al. (1999), are shown. Hillenbrand et al. report mean values across multiple child (10–12 years) and adult female talkers; note that diphthong pitch and formant information were not available in that dataset. Relevant values for 7-year-old children and adult female talkers were selected from Lee et al. for the vowels used in the present study.

Table 2 shows the ratio of frequency values between the child and adult talkers, calculated for the present stimuli as well as for vowel productions reported in Hillenbrand et al. (1995) and Lee et al. (1999). The child-to-adult frequency ratios were broadly comparable across datasets. Across all stimuli, the mean ratio between the child and adult talkers was 1.15 ± 0.18 for the present stimuli, 1.08 ± 0.04 for Hillenbrand et al., and 1.18 ± 0.07 for Lee et al.

TABLE 2.

Top: Ratios of mean values for f₀, F1, F2, and F3 values between the present study and Hillenbrand et al. (1995) and Lee et al. (1999) for female child and adult talkers

Present Study		Hillenbrand et al. (1995)				Lee et al. (1999)
		f 0	F1	F2	F3	f 0	F1	F2	F3
Child	f 0	1.05				1.04
	F1		1.03				1.00
	F2			1.03				1.14
	F3				0.88				1.15
Adult	f 0	0.97				0.93
	F1		0.88				1.05
	F2			0.99				0.98
	F3				0.93				0.96

	Child	Adult
		f 0	F1	F2	F3
Present study	f 0	1.18
	F1		1.23
	F2			1.19
	F3				1.01
Hillenbrand et al. (1995)	f 0	1.08
	F1		1.04
	F2			1.11
	F3				1.09
Lee et al. (1999)	f 0	1.20
	F1		1.20
	F2			1.17
	F3				1.14

Data for child talkers were from 10-to-12-year-olds in Hillenbrand et al. (1995) and from 7-year-olds in Lee et al. (1999). Bottom: Mean ratios for f₀, F1, F2, and F3 between female child and adult talkers from the present study, Hillenbrand et al., and Lee et al.

Table 3 shows the consonant–vowel (CV) ratios (top), the mid-frequency energy (middle), and the spectral tilt (bottom) for the present digits produced by a female child and adult. The mean CV ratio across all digits was −18.1 ± 9.5 dB for the child talker and −15.3 ± 8.8 dB for the adult talker. The mid-frequency energy was calculated as in Ferguson et al. (2024). The mean mid-frequency energy across all digits was 30.8 ± 11.8 dB for the child talker and 39.9 ± 9.8 dB for the adult talker. The adult mid-frequency energy was comparable to that reported in Ferguson et al., where the mean mid-frequency energy was 37.7 ± 0.2 dB. Spectral tilt was calculated H1 – A3, as in Iseli et al. (2007). The mean spectral tilt across the digits analyzed was 4.5 ± 1.9 dB for the child talker and 8.3 ± 2.4 dB for the adult talker. Note that while the present mean spectral tilt values are lower than reported in Iseli et al., the difference in spectral tilt values between the female child and adult talkers is comparable (approximately 4 dB). One issue that makes it difficult to compare CV ratios, mid-frequency energy, and spectral tilt across studies is the different stimuli used in each study to extract speech features. In the present study, features were extracted from digits where there were transitions between different consonants and vowels within a speech token. This is quite different from the h/V/d and v/C/v vowels in Hillenbrand et al. (1995), the “b”/V/c in Lee et al. (1999), or the synthetic vowels in Iseli et al.. Still, speech features from the present female child and adult talkers are generally comparable to those from previous studies, despite differences in stimuli.

TABLE 3.

Consonant–vowel ratios (top), mid-frequency energy (middle), and spectral tilt (bottom) for the present digits produced by a female child and adult

Consonant–Vowel Ratio (dB)
Digit	0 (“z,” “o”)	1	2	3	4	5	6	7	8	9	Mean	SD
Child	−12.4	−5.1	−11.9	−15.9	−32.0	−28.1	−28.0	−24.0	−17.1	−6.2	−18.1	9.5
Adult	−14.7	−3.9	−10.9	−3.6	−25.0	−20.4	−16.5	−22.9	−28.0	−7.5	−15.3	8.8
Mid-Frequency Energy (dB)
Digit	0	1	2	3	4	5	6	7	8	9	Mean	SD
Child	36.8	37.7	36.8	51.9	22.6	11.6	19.1	25.1	39.4	27.2	30.8	11.8
Adult	43.3	53.9	42.0	44.6	46.6	18.0	31.5	44.7	37.7	36.3	39.9	9.8
Spectral tilt (dB)
Digit	0 (“o”)	1	2	3	4	5	6	7	8	9	Mean	SD
Child	6.9	3.6	5.8	4.9	8.4	3.0	−2.9	3.2	2.3	3.0	4.5	1.9
Adult	8.7	7.9	9.9	7.4	13.5	4.2	6.1	8.5	6.5	10.0	8.3	2.4

dB, decibel.

To account for the potentially wide range of age-related SRTs, a fixed overall presentation level was used instead of a fixed target or masker level to prevent excessive loudness for some listeners. After combining the target (three-digit sequence) and SSN maskers at the designated SNR, the overall level was adjusted to ensure a consistent root-mean-square amplitude across all trials and test runs. All stimuli were presented at 65 dBA via a loudspeaker (8030C, Genelec Oy, Iisalmi, Finland) placed 1 meter in front of the listeners’ heads. During each test trial, target digits were presented at the desired SNR; the initial SNR was 10 dB. Participants were instructed to listen to the target digits and click on response buttons (labeled 0, 1, 2, 3, 4, 5, 6, 7, 8, and 9) in the order of the digits heard. If all three digits were correctly identified, the SNR was reduced by 4 dB (initial step size) by increasing the level of the SSN masker; if not, the SNR was increased by 4 dB. After two reversals in SNR, the step size was reduced to 2 dB. The SRT was calculated by averaging the last six reversals in SNR. If there were fewer than six reversals within 25 trials, the test run was discarded, and another run was measured. Two test runs were completed for each listening condition, and the SRT was averaged across runs. The target conditions and multiple runs were randomized within and across listeners.

Statistics

The effect of the target talker (adult versus child) between the two listener groups (children and adults) was analyzed using a repeated measures analysis of variance (RM-ANOVA). To examine the developmental effects among children, a linear mixed-effects model (LMEM) was used to examine the relationship between SRT, age (log-transformed), and target talker (adult versus child), while accounting for subject-level variability with a random intercept. The model included fixed effects for these predictors, as well as two-way and three-way interactions (Age × Target Talker, Sex × Target Talker, Age × Sex, and Age × Sex × Target Talker).

RESULTS

All children and adults completed all SRT measurements. Table 4 shows mean SRTs with the adult and child-target talkers for different age groups of children and for adults; mean SRTs from Koopmans et al. (2018) groups are shown for comparison. Figure 2A shows violin plots of SRTs for adult and child groups listening to adult and child-target speech. Children showed mean SRTs of −8.07 dB (SD = 1.94) for the adult-target talker and −9.49 dB (SD = 1.22) for the child-target talker, whereas adults showed lower (better) SRTs of −9.62 dB (SD = 0.99) and −10.6 dB (SD = 0.97) for the same conditions. Note that a potential outlier, the 4-year-old’s SRT for the adult-target talker, was removed. Residuals were computed by regressing out log-transformed age from both adult- and child-talker SRTs for children. A Pitman–Morgan test on these residuals indicated significantly greater age-adjusted variance for the adult than the child-talker condition, F_{(68, 68)} = 2.37, p < 0.001 (ratio of variances = 2.36, 95% confidence interval: 1.47 to 3.83). Figure 2B plots the residual SRTs for both target types for children.

TABLE 4.

Mean SRTs in dB SNR for various age groups

Group	N	Adult Target	Child Target
4–5 yrs	6	−4.05* (1.88) [−5.8]	−8.10 (0.44)
6–8 yrs	21	−7.38 (0.40) [−7.4]	−9.28 (0.34)
9–12 yrs	13	−8.61 (0.49) [−8.4]	−9.90 (0.24)
13–17 yrs	30	−8.72 (0.26)	−9.73 (0.16)
Children	70	−7.90 (0.29)	−9.49 (0.15)
Adults	27	−9.62 (0.19) [−9.5]	−10.59 (0.19)

Values in parentheses and brackets are the standard error of the mean and mean values from Koopmans et al. (2018), respectively.

^*For the 4–5-year-old group, the mean was −5.6 dB after removing a questionable outlier.

Fig. 2.

Speech recognition thresholds across listener groups and talker types. A, Violin-box plots for speech recognition thresholds across children and adults for both adult and child targets. The violins show the kernel density distribution for each group and targets, with wider sections indicating higher data density. The box shows the IQR, with the median indicated by a horizontal bar within the box. The whiskers show 1.5 times the IQR. B, Violin-box plots of age-adjusted speech recognition threshold residuals for both adult- and child-target talkers for children. IQR indicates interquartile range.

An RM ANOVA was conducted on the SRT data, with target talker (adult, child), listener group (adult, child), and listener sex (male, female) as factors. Results showed significant main effects for target talker [F(1,93) = 35.05, p < 0.001, η_p² = 0.27] and listener group [F(1,93) = 18.63, p < 0.001, η_p² = 0.17], but not for listener sex [F(1,93) = 0.04, p = 0.847]. No significant interaction was observed between the target talker and listener group [F(1,93) = 1.02, p = 0.316], suggesting that the effect of the target talker did not significantly differ between the listener groups. No significant interactions were observed between listener sex and target talker [F(1,93) = 0.07, p = 0.785] or listener group [F(1,93) = 0.002, p = 0.967], or among listener sex, target talker, and listener group [F(1,93) = 3.23, p = 0.076]. Excluding the outlier data from the 4-year-old did not affect the RM-ANOVA outcomes; effect sizes increased only marginally.

Figure 3A shows a scatterplot of SRTs for the child listening group as a function of age. LMEM demonstrated a good fit, with a conditional R² = 0.52 (indicating total variance explained, including random effects) and a marginal R² = 0.36 (explaining variance due to fixed effects alone). The model significantly improved fit compared with a null model [χ²(4) = 68.83, p < 0.001]. Age significantly affected SRTs (β = −5.71, SE = 1.062, p < 0.001), with SRTs improving with age. There was no effect of listener sex (β = −0.03, SE = 0.32, p = 0.92). Target talker was a significant predictor (β = −1.60, SE = 0.245, p < 0.001), with significantly lower (better) SRTs with the child than with the adult talker. There was a significant interaction between age and target talker (β = 4.99, SE = 1.642, p = 0.003). There were no significant interactions between listener sex and target talker (β = −0.61, SE = 0.49, p = 0.22) or age (β = 0.66, SE = 2.12, p = 0.759), or among age, listener sex, and target talker (β = −0.95, SE = 3.28, p = 0.77). When the outlier data point for a 4-year-old listening to an adult talker (SRT = 3.7 dB) was removed, LMEM analysis showed similar results, with the significant interaction between age and target talker persisting (β = 3.06, SE = 1.44, p = 0.038). For this model, the semi-partial R² values indicate the unique contributions of each predictor to the explained variance in the SRT. Age had a moderate effect (semi-partial R² = 0.21), while the talker type (adult versus child) showed a large effect (0.28). The interaction component (Age × Talker Type) had a small effect (0.10).

Fig. 3.

Age-related changes in speech recognition thresholds. A, SRTs as a function of age for children, adults, and child targets, with box-whisker plots for adult data (broken bars as mean). Lines show linear fits with 95% CIs. B, Age-SRT function for the adult target with a significant breakpoint at 10.47 years (1.02 in log units). The horizontal line shows the 95% CIs for the breakpoint. Data are not shown for the child target because the breakpoint was not significant; as such, data for the child target are shown in A. CI indicates confidence interval; SRT, speech recognition threshold.

To further examine the significant interaction between age and target talker, two segmented regression analyses were performed to determine breakpoints in SRTs versus age for adult and child targets. To obtain robust estimates of breakpoints, data from both adult and child listeners were considered together. For the adult target, a significant breakpoint was observed at an age of 5.4 years (log age = 0.73), suggesting that the relationship between age and SRT changes at this point. Before the breakpoint, the slope significantly decreases (p = 0.004), and after the breakpoint, the slope shifts by +40.52. That is, the slope becomes −4.37 (calculated as −44.89 + 40.52), indicating that the change in SRT is less steep compared with before the breakpoint. Segmented regression analyses, performed after removing the potential outlier (4-year-old; 3.7 dB), revealed a significant breakpoint at an age of 10.5 years (log age = 1.02) with a steep slope (p = 0.004). After the breakpoint, the slope shifts by +6.71, meaning the slope becomes −2.23 (calculated as −8.94 + 6.71), indicating that the change in SRT for the adult target is less steep after the breakpoint compared with before it. For the child target, a breakpoint was detected at 5.9 years (log age = 0.77), but the slope was not statistically significant (p = 0.46). In addition, there was no significant effect of sex of the listener or interaction between sex and age of the listener in any of the segmented regression analyses (p > 0.05 in all analyses), meaning that the breakpoints did not change between male and female listeners. Given that the breakpoint was not significant for the child target, Figure 3B shows the breakpoint in the age versus SRT function for the adult target.

For the child target, analysis of variance on ranked SRT data was performed to determine the age at which there was no significant difference between children and adults. Listener data were grouped into 4 to 5, 6 to 7, 8 to 9, 10 to 11, 12 to 13, 14 to 15, 16 to 17, and >18 years old. There was a significant effect of age group [H(7) = 23.31, p < 0.001]; post hoc Dunn’s tests showed SRTs were significantly different between adults and children aged 4 to 5 and 6 to 7 years (p < 0.05), with no significant difference among the remaining age groups. This may suggest that child SRTs became adult-like by age 8 for the child target.

A partial correlation analysis was conducted to examine the relationship between SRTs for adult and child target speech while controlling for age effects. Results revealed a significant positive correlation between adult and child-target speech (Pearson’s r = 0.25, two-tailed p = 0.014, n = 97), indicating a small effect size; the coefficient increased to 0.46 after removing the outlier data from a 4-year-old (p < 0.001). Figure 4 shows the scatterplot with the SRT for the child target plotted against the SRT for the adult target.

Fig. 4.

Speech recognition thresholds for the child target as a function of the adult target for child and adult listeners. The trendline is age-adjusted [(SRT for the child target = −6.2 + 0.145 × (SRT for adult target) − 2.2 × (Log Age)]. SRT indicates speech recognition threshold.

DISCUSSION

Main Findings

The present study represents an initial investigation into differences in speech recognition in noise by children and adults listening to target speech produced by child- or adult-talkers. Speech perception assessment in children has predominantly utilized adult talkers as the target, potentially overlooking key developmental differences in processing speech from peer-aged talkers. Contrary to our hypothesis, results showed that SRTs in child listeners were lower (better) with the child target than with the adult target. It is interesting that the effect of the target talker (adult or child) did not differ significantly between child and adult listeners, suggesting that both listener groups exhibited similar responses to adult and child talkers. However, a significant interaction was observed in child listeners between age at testing and target talker. While there was a greater overall advantage with the child than with the adult target, age effects were more pronounced for the adult target. Moreover, the weak correlation between SRTs for child and adult targets underscores the limitations of relying solely on speech perception tests with adult talkers when evaluating perceptual abilities in children. While these initial findings may be of great interest, they may also be driven by the specific talkers and speech materials (digits) used in the present study. Further testing with a larger number of adult and child talkers and/or alternative speech materials may elucidate differences in perception of speech produced by adults and children.

Talker Age Effects

The present study highlights the influence of talker age on speech-in-noise recognition, consistent with Hazan and Markham (2004), but it also reveals important distinctions. Hazan and Markham reported that word recognition scores were lower for child targets (approximately 13-year-olds) compared with adult female targets for younger children (7 to 8 years), older children (11 to 12 years), and adults. In contrast, the present study found that both adult and child listeners achieved significantly better SRTs for a child female target (7 years) compared with an adult female talker. Besides the age of the talkers and differences in scoring methods, another important difference was that Hazan and Markham’s study used samples (words) from a group of male and female talkers producing naturalistic speech, while the present study utilized clear speech produced by a single female child talker. In addition, in Hazan and Markham, the profiles of the “best” talkers did not include any child talkers, while the profiles of the “worst” talkers included a male child. These differences may have contributed to the apparent disagreement in findings. However, both studies consistently found that the pattern of recognition (whether better or poorer performance for child-target speech) did not differ between child and adult listeners, indicating that target talker-related differences in speech-in-noise recognition might reflect similar underlying mechanisms in both groups.

An interesting finding was the clear difference in variability between SRTs for adult and child target talkers. The larger variance (standard deviation) observed for the adult target is consistent with previous reports of inter-subject variability when both children and adults are tested with adult talkers in noise (Koopmans et al. 2018; (Van den Borre et al. 2021; Mishra et al. 2022). In contrast, the markedly reduced variance for the child target indicates that listeners were more homogeneous in their performance when processing child-produced speech, potentially reflecting more consistent perceptual strategies or reduced acoustic–phonetic variability in how the child talker was encoded. Together, these results demonstrate that variability in speech-in-noise performance is not uniform across target talkers: adult talkers elicit a broader spread in SRTs documented in prior literature, whereas child talkers yield more tightly clustered perceptual outcomes.

Developmental Differences

Developmental effects in child listeners differed between the child and adult target talkers, with a steeper slope observed for the adult than for the child talkers (Fig. 3). The use of the different SSN maskers for the adult and child talkers may have sharpened the differences in these slopes, as there was maximal spectral overlap between the masker and target speech. Had a single masker been used, age effects may have been diluted due to the incomplete masking of the adult or child talkers. It is interesting that the breakpoint at 10.5 years in the age-SRT function for the adult target was remarkably similar to the 10.2 years observed in a previous study using an adult male target under headphones (Mishra et al. 2022). However, no significant breakpoint was observed for the child target. This suggests that developmental effects may depend on the age of the talker. For child listeners ≤10 years old, SRTs were markedly better with the child than with the adult target talker. This may be due to the greater similarity in speech patterns between the child listeners and the child target, and the greater dissimilarity between the child listeners and the adult target. As such, the internal speech pattern templates within the younger listeners may have contributed to the better SRTs with the child target. According to the theory of internal speech patterns, individuals use stored representations of speech sounds, which are shaped by their own experiences with speech input, to facilitate the recognition and processing of speech (Oppenheim & Dell 2010; Alderson-Day & Fernyhough 2015; Fernyhough & Borghi 2023). In younger listeners, these internal templates may have been more closely aligned with the speech patterns of the child target, allowing for more efficient processing and better speech recognition. The internal speech patterns may have become more similar between older children (≥11 years old) and the adult target. Developmental effects may have allowed older children to accommodate the speech variability in the child target as well.

Implications

Study findings have several important implications. First, in speech perception research, the results highlight the need to consider child-produced speech as a relevant and distinct factor in evaluating how children process speech. Traditional models of speech recognition often rely on adult talkers, potentially overlooking differences in how children perceive speech in real-world environments where peer talkers are more common. These findings suggest that the development of speech recognition abilities in children may not follow a uniform pattern for adult versus child targets, which could influence how speech perception is studied across age groups. Second, from an auditory development perspective, the results point to the evolving nature of speech recognition as children age. The age-related differences in performance suggest that younger children may benefit more from exposure to child talkers, possibly due to their unique familiarity with peer speech, which may support both speech recognition and the development of speech production. As children mature, there may be a developmental shift toward adult talkers serving as more relevant models. This could reflect changes in auditory processing and cognitive development, as children become more adept at handling more complex auditory inputs. Third, the differences in the perception of child and adult targets highlight the relevance of studying the perceptual implications of the development of control over voicing parameters such as pitch, loudness, and sound duration. Finally, the clinical implications of these results highlight the need for more personalized approaches when assessing speech perception in noisy environments. Given that children process speech from both child and adult talkers differently, hearing devices should be programmed with careful consideration of how children perceive peer speech, particularly in challenging listening environments. These findings also underscore the importance of incorporating child-target speech stimuli in evaluation tools. This would not only enhance the ecological validity of auditory assessments but also ensure more effective fitting and rehabilitation of hearing devices for children.

Limitations

The interpretation of the present data is limited because of the use of digits and single talkers as target stimuli; a greater number of target talkers and alternative target speech stimuli may provide detailed insight into differences in perception of speech produced by adults or children. The primary motivation for using digits was to reduce the influence of cognitive factors in young child listeners (as young as 4 years), because sentence recognition is more susceptible to such effects (Sanchez et al. 2022). While this approach helps minimize variability related to language and cognition, it may also limit the generalizability of the findings. In addition, broader acoustic variability likely exists among female child and adult talkers beyond the single talkers selected for this study, which may further constrain generalizability. However, as shown in Tables 1 and 2, the fundamental frequency and formant values observed in the present recordings are comparable to those reported in the literature (Hillenbrand et al. 1995; Lee et al. 1999), suggesting that the selected talkers were reasonably representative of typical female speech in children and adults. Still, further testing with more child and adult talkers is needed to fully understand differences in perception of speech produced by adults and children.

CONCLUSIONS

The present study provides evidence that speech recognition may differ significantly between adult and child-target talkers, particularly in the presence of noise. Children performed better with a child target than with an adult target, and younger listeners were most affected when listening to adult talkers. Performance variability was greater for adult targets and more homogeneous for child targets. While many questions about speech perception across talker- and listener-age groups remain to be answered, this study provides important initial insights into how target talker age shapes both the mean and variability of speech-in-noise performance, highlighting the need to include child-produced speech in research and clinical assessments.