Auditory-Perceptual Features of Speech in Children and Adults With Down Syndrome: A Speech Profile Analysis

Abstract

Purpose

The aim of this study was to determine how the speech disorder profiles in Down syndrome (DS) relate to reduced intelligibility, atypical overall quality, and impairments in the subsystems of speech production (phonation, articulation, resonance, and prosody).

Method

Auditory-perceptual ratings of intelligibility, overall quality, and features associated with the subsystems of speech production were obtained from recordings of 79 children and adults with DS. Ratings were made for sustained vowels (62 of 79 speakers) and short sentences (79 speakers). The data were analyzed to determine the severity of the affected features in each speaking task and to detect patterns in the group data by means of principal components analysis.

Results

Reduced intelligibility was noted in 90% of the speakers, and atypical overall speech quality was noted in 100%. Affected speech features were distributed across the speech production subsystems. Principal components analysis revealed four components each for the vowel and sentence tasks, showing that individuals with DS are not homogeneous in the features of their speech disorder.

Discussion

The speech disorder in DS is complex in its perceptual features and reflects impairments across the subsystems of speech production, but the pattern is not uniform across individuals, indicating that attention must be given to individual variation in designing treatments.

Down syndrome (DS) is a genetic defect caused by trisomy of whole or part of chromosome 21. Worldwide, it is the most prevalent genetic disorder and the most common genetic cause of intellectual disabilities (Kazemi et al., 2016). DS often has multiple effects on body systems and organs (Roizen, 2010), but life expectancy for individuals with DS has increased from about 30 years in the 1980s to about 60 years in 2020 (Bull, 2020). This increased life expectancy offers expanded opportunities for education, recreation, employment, and social interaction. One of the obstacles to improved quality of life is that individuals with DS often have atypical patterns of speech and language that can interfere with communication and social interaction (Roberts et al., 2007). Although difficulties with communication often occur in individuals with intellectual disability, it has been reported that people with DS are 2.6 times more likely to have moderate communication difficulties and 1.9 times more likely to have severe communication difficulties than people with intellectual disability not associated with this syndrome (Smith et al., 2020). Speech is the main form of communication in 97% of all people with DS, making it central to considerations of personal interaction (Coppens-Hofman et al., 2012). It has been suggested that speech impairments could serve as flags to identify children with DS who are at risk for adaptive behavior difficulties (Stephan et al., 2021).

Speech and voice in DS are affected by a complex of disturbances that can result in reduced speech intelligibility and atypical quality (Kent & Vorperian, 2013; Kumin, 1994; Wild et al., 2018; E. M. Wilson et al., 2019b). Reduced intelligibility can have profound negative consequences, such as being a barrier to educational achievement (van Bysterveldt et al., 2019), influencing peer victimization of adolescents with DS (Reardanz et al., 2020), and limiting the effective use of voice-enabled technologies such as speech recognition (e.g., Pascu, 2019; Project Understood, n.d.). The complexity of speech production characteristics in DS is purportedly the result of any combination of causes, including anatomic differences (Korayem et al., 2019; Rodrigues et al., 2019; Sforza et al., 2012), motor speech disorders (Coêlho et al., 2020; Kumin, 2006; Rupela et al., 2016; E. M. Wilson et al., 2019a), cognitive–linguistic disorders (Carr, 2012; Lukowski et al., 2019), sensorimotor disorders (Weeks et al., 2000), hearing loss (Nightengale et al., 2017; Porter & Tharpe, 2010), and comorbid conditions, such as autism spectrum disorder (Capone et al., 2005) and/or attention-deficit/hyperactivity disorder (Ekstein et al., 2011).

Given the potential multifactorial causative nature of speech and voice production difficulties in DS, it is predictable that deficits may occur across all speech subsystems, for example, respiratory, phonatory, articulatory, resonatory, and prosodic. A subsystem approach to speech analysis has precedent in the study of complex disorders such as dysarthria in children (Chen et al., 2018; J. Lee et al., 2014) and adults (De Bodt et al., 2002; Lowit & Kent, 2017; Schölderle et al., 2016). The details of subsystem description vary across published reports (see examples in Chen et al., 2018), but the common goal of studies using this approach is to document how the various components of the speech production process are affected in a speech disorder. Recent studies reinforce the conclusion that the speech disorder in DS derives from a combination of impairments in individual subsystems. In an auditory-perceptual study of 26 children with DS (ages 8 years and older), Jones et al. (2019) determined that “speech disorders in DS reflect distributed speech production impairments involving voice, speech sound production, fluency, resonance, and prosody” (p. 330). Similarly, E. M. Wilson et al. (2019a, 2019b) reported across-subsystems deficits in their studies involving 45 children and adolescents with DS and concluded that “lowered intelligibility was significantly associated with across-the-board phonemic and phonetic errors and inappropriate prosody and voice” (E. M. Wilson et al., 2019b, p. 810).

The multidimensional nature of speech and voice production features in individuals with DS presents a clinical challenge for assessment and treatment (Meyer et al., 2017), but further complicating this effort are the varying and often discrepant descriptions across studies. For example, phonatory dysfunction in DS has been variously described as vocal hyperfunction (Pebbili et al., 2019), vocal hypofunction (Moran & Gilbert, 1982; Wold & Montague, 1979), or nonmodal phonation (Jeffery et al., 2018). Atypical resonance, which is commonly attributed to the speech of persons with DS, can take different forms, including hypernasality (Rolfe et al., 1979), hyponasality (Jones et al., 2019), or other atypical resonance (Fourakis et al., 2010; Jones et al., 2019). Prosody is often disturbed in the speech of individuals with DS, but studies indicate different patterns, such as general deficits in production and not perception (Stojanovik, 2011), or specific abnormalities (e.g., deficits in interrogative intonation) but a basic similarity to healthy controls (Zampini et al., 2016). It is also possible that the dysfunctions in a given individual vary across observations or even within an utterance (e.g., alternations between hypernasality and hyponasality, or between breathiness and roughness). As reviewed in Kent and Vorperian (2013), published reports differ on several other speech and voice characteristics, including speaking rate, pitch level, and types of disfluency. Given the discordant conclusions, it is difficult to reach a confident generalization about speech features in DS even though there is consensus that speech production is commonly affected and interferes with effective oral communication.

The importance of highlighting these differences across investigations is not to suggest erroneous modeling, but rather that this variance is likely a reflection of the heterogeneity of speech and voice production characteristics within and across individuals as well as the methodological challenges related to comprehensively quantifying complex, multilayered disorders (Björelius & Tükel, 2017; Kent, 2000). Yet, it is imperative to recognize the lack of consensus on these basic aspects of speech production in DS because it impedes the design of clinical interventions and the selection of outcome measures for these interventions. For example, if DS is associated with phonatory hypofunction, then an intervention such as the Lee Silverman Voice Treatment might be indicated. However, if phonatory hyperfunction is involved, then the Lee Silverman Voice Treatment would not be appropriate and could even be contraindicated.

The research reported here complements previous studies in three major respects. First, we report data on speech production for both children and adults, covering the age range of 4–55 years. Because most published studies on speech in DS pertain to either children or adults, but not both, there is limited information on how speech production may change over the life span in persons with DS. A life span perspective is particularly warranted given the increased life expectancy rates in persons with DS (Bull, 2020). Second, we do not assume homogeneity in the speech characteristics in DS but rather test the hypothesis that there are subgroups of individuals with DS, each with distinctive patterns of speech impairment. An advantage to this approach is that it can contend with factors that complicate research in DS, including phenotypic variation (Fidler, 2005), sex differences (Wild et al., 2018), and individual differences (Karmiloff-Smith et al., 2016; Thomas et al., 2020). Third, because the speech impairments in DS appear to be distributed across the components of the speech production system, we adopt a speech subsystems approach in reporting and interpreting the data. The speech subsystems perspective is coupled with a task-based analysis in which we collect data for different speaking tasks. A given task makes specific requirements of the speech subsystems (Green et al., 2013; Kent & Kent, 2000). For example, vowel prolongation relies primarily on the respiratory–phonatory subsystem and makes minimal demands on supralaryngeal articulation other than maintenance of an open airway. In contrast, production of isolated words requires coordination of supralaryngeal articulatory movements with respiration and phonation, and production of sentences goes further to involve prosodic features. Data for single words were reported previously from our group (Wild et al., 2018), and this present report focuses on vowel prolongation and sentence production. A major advantage of this task-based approach is that it can address variations in severity of disorder across the subsystems of speech production, thereby leading to a more comprehensive speech profile analysis of the speech disorder.

Therefore, the goal of this project is to determine if there are distinct speech profiles in DS and, if there are, how these profiles relate to subsystem impairments and degree of intelligibility. The method used is auditory-perceptual rating of multiple features that relate to the subsystems of speech production. Auditory-perceptual judgments are one of the primary tools used by speech-language clinicians to assess speech production and to evaluate the effects of intervention. The determination of speech profile(s) and subsystem involvement based on this approach would provide the foundation for the future development of profile-specific evidence-based assessment and treatment protocols to most proficiently improve communicative effectiveness in individuals with DS.

Method

Auditory-perceptual ratings of selected features of speech production were the source of data in this study. Speech materials, speakers, and raters are described below. This research was approved by the Health Sciences Institutional Review Board of the University of Wisconsin-Madison.

Speech Production Protocol

The speech materials used in this study were selected from the acoustics database of the Vocal Tract Development Lab where the same speech production protocol was used across all ages from speakers with DS and typically developing speakers. The speech production protocol in the larger study included high-quality recordings of the following speaking tasks: sustained production of four vowels /i u ɑ ae/, a set of 20 isolated monosyllabic words, and six short sentences or phrases. For the purposes of this perceptual rating study, the following stimuli were selected from speakers with DS: sustained production of vowel /ɑ/ and repetition of three brief sentences (Cars go beep beep, Pop the bubble, and The blue duck quacks). These sentences were chosen for their phonetic variety and likelihood of production by the youngest speakers.

Participants

This report is based on a subset of the original recordings in the databases where only the initial/first visit recordings from 79 separate/individual participants with DS, ages 4–55 years, were used, that is, repeat longitudinal recordings were excluded. Available vowel recordings were from 62 speakers with DS (32 males, 30 females), and sentence recordings were from 79 speakers with DS (41 males, 38 females). The number of speakers/recordings for the two tasks was different because nine participants did not complete the vowel production task, and eight vowel recordings were eliminated because they were deemed inappropriate for analysis due to either a “purposeful vibrato” or the vowel prolongation was too short for auditory rating.

Raters

The three raters had substantial listening experience in judging various types of disordered speech and specifically with speech produced by children and adults with DS. To prepare for the listening tasks, the raters discussed the features to be rated and agreed on the definitions given in the Appendix. The raters agreed on grouping features into the main categories of phonation and respiration, supralaryngeal articulation and resonance, suprasegmental features of speech, and overall communication effectiveness. These categories relate to the subsystem perspective that is used to describe atypical function at various levels and also relate to global aspects of speech production. The raters then participated in consensus training. They collectively listened to 13 selected vowel /ɑ/ samples (six female, seven males; age range: 5–33 years) and 17 speaker's sentence productions (seven females, 10 males; age range: 5–33 years). After the individual ratings were completed for a given sample and speaker, the raters compared their results and discussed possible reasons for differences in ratings. In doing so, the raters refined feature definitions as needed and attempted to identify factors that could have led to different ratings for a given feature (e.g., differences related to speaker age and/or sex). In most cases, the differences were in rating the degree of severity of an affected feature.

Stimuli

The vowel /ɑ/ prolongation and sentences recordings from speakers with DS, tasks in a larger speech production protocol, were segmented manually based on criteria described by Wild et al. (2018) using Praat 5.1.3a and saved as separate sound files. These files were normalized to −10 dB amplitude using Sound Forge 8.0 to ensure that fluctuations in the sound level would not affect perceptual ratings. Listening sessions consisted of the presentation of the vowel stimuli first, followed by the sentence stimuli. The three sentences were presented consecutively, in the same order, and rated together as a unit. This was done to provide the raters with a sample of sufficient length and phonetic variation. The stimuli were organized by the speakers' identification code and age at visit to correspond with the groupings in Wild et al. to ensure the balanced distribution of age and speaker sex. In addition, to assess intrarater reliability, a subset of the vowels and sentences was selected using Excel's random number generator (Microsoft Office Professional Plus 2013, Excel, 64 bit). The vowel reliability set consisted of 21 speakers (10 males, 11 females), and the sentence reliability consisted of 25 speakers (16 males, nine females). Reliability outcome data are reported in the Statistical Analysis section below.

Rating Procedure

The rating scale selected for this study was a hybrid ordinal scale that begins with a category denoting absence of a particular feature (e.g., “not present”), followed by levels of ordered presence (e.g., severity) of the phenomenon. Cicchetti (1976) termed this scale as dichotomous-ordinal scale, and its clinical application was discussed further by Cicchetti et al. (2006). Such a scale of presence/severity is well suited to assessments of disorders involving several features of potential abnormality. When several features are not remarkable (i.e., they are not judged to be present) in perceptual judgment, indices of interrater agreement can be inflated by their inclusion. As Kearns and Simmons (1988) pointed out, estimates of reliability in perceptual ratings can be spuriously inflated when only a small percentage of the perceptual features are perceived as disordered (i.e., present). The features (as listed in the Appendix), except “speaking rate” and “pitch level,” were judged on a 5-point interval scale, where 1 indicates that the feature is not present and numerals 2 through 5 indicate degree of severity (2 = mild, 3 = moderate, 4 = severe, 5 = very severe). The bipolar rating for speaking rate and pitch level was 1 for too slow or too low and 5 for too fast or too high, respectively, for a given age/gender. The Appendix lists the features within the various subsystems and overall communication effectiveness.

An in-house computer program was written using Python 3.2 for stimuli presentation and scoring of ratings. The three raters independently performed the listening tasks in a quiet room, with the stimuli presented through dynamic stereo headphones, with frequency ranging from 18 Hz to 18 kHz. The computer program displayed a single screen for each speaker with DS during stimulus presentation. The screen for both vowels and sentences listed gender and age at visit, along with a “play button” to initiate or repeat stimulus presentation and a “next button” to move to the following speaker. The vowel screen displayed 13 speech features for rating, and the sentences screen displayed 22 features. See the Appendix for a listing of the features used in rating both tasks. Bipolar features, along with a space for typing comments, were located at the bottom of the respective screens. The raters could choose to hear each vowel presentation up to 3 times and the sentence stimuli up to 4 times. The program would not advance to the next speaker until all feature/attribute fields were rated. Perceptual ratings were done across several sessions, owing to the large number of speakers and stimuli to be rated. Typically, the rating sessions were done for no longer than 1 hr at a time to avoid fatigue or loss of attention. The overall process occurred over several weeks, with a different number of sessions for each rater. The mean durations for the vowel ratings for each sample by the three raters were 89, 85, and 49 s. Mean durations for the sentence ratings were 172, 228, and 111 s. The computer program compiled and tabulated all scores into .txt documents coded with the raters' initials for analysis. After the completion of the vowels and sentence ratings, each listener rated a subset of selected samples to assess intrarater reliability.

Statistical Analysis

In the following sections, we report on a series of reliability, principal component, and correlational analyses. All analyses were performed using SPSS Version 25.

Reliability: Intrarater and Interrater

To assess consistency in feature scoring, both intrarater (within, based on ratings by the same listener of the same speech segment at two time points) and interrater (between, based on ratings by different listeners of the same speech segment) reliabilities were assessed. Formal assessment using Cohen's kappa (Cohen, 1960) was performed. Kappa coefficients were averaged for individual raters across features in evaluating intrarater reliability. Results confirmed that raters had better intrarater reliability for the sentences (κ = .27–.33), with each listener/rater having fair reliability, than for vowels (κ = .09–.27), Furthermore, findings revealed feature-specific variability in intrarater reliability, where vowel features, such as roughness and pitch variability, and sentences features, such as roughness, breathiness, dysphonia severity, distorted vowels, and reduced intelligibility, had fair-to-moderate kappa coefficients (κ = .21–.61) across all three raters, while other features, such as pitch breaks and the subtypes of atypical resonance (listed in the Appendix), were not reliable. Similar to the results for intrarater reliability, interrater reliability was generally better for sentences than vowels. However, overall reliability tended to be slight to fair and again showed feature variability. For example, the features breathiness and loudness variability for vowels and imprecise consonant articulation, reduced intelligibility, atypical overall quality of speech, disfluency, and speaking rate had fair kappa coefficients (κ = .21–.40), while the rating of features such as the subtypes of atypical resonance were not reliable. Thus, with the exception of the reduced intelligibility feature, both intra- and interrater reliability kappa values ranged considerably across features and lacked desirable levels of agreement (moderate to substantial kappa values: κ = .41–.60 and κ = .61–.80, respectively; see the Discussion section for comments on the problem of low reliability for both individual raters and features that has plagued auditory-perceptual studies of speech and voice). To help address this problem, we sought to define rater/feature composites, as opposed to individual feature ratings, in our subsequent analyses. To this end, we both (a) constructed a rule-based score for each feature across raters and (b) applied a dimensionality reduction strategy to define meaningful composites of the features as opposed to analyzing each feature separately. We describe each of these approaches in the following section.

Dimensionality Reduction of the Feature Rule-Based Scores and Relationships to Speech Intelligibility

To address the less than optimal/desired levels of intra- and interreliability at the individual rater and feature levels, we first applied a rule-based scoring strategy that incorporates ratings from multiple raters in scoring each feature. This strategy is a modification of the approach of odd-rater out. The rule-based strategy was as follows:

1. If two out of three listeners had the same rating, that common value was taken as the consensus score.

2. If the ratings fell into a consecutive range (e.g., 2, 3, 4), the median value was taken as the consensus score. However, if the series contained a score of 1 (meaning the feature was not present), the value of 1 was excluded and the average of the other two ratings was taken as the consensus score. For example, with the series (1, 2, 3), the average is 2.5. In this way, only the ratings of severity among raters who actually recognized it are taken into account.

This rule-based scoring strategy for each speaker with DS highlights perceptual attributes that were recognized by the raters and the degree of severity when a feature was considered to be present. The frequency of use of Rule 1 versus Rule 2 is shown in Table 1. The vowel data demonstrated agreement of raters (Rule 1) at least 70% of the time. The sentence data demonstrated exact agreement between at least two of raters (Rule 1) 75% of the time for 15 features, with the remaining three features (roughness, dysphonia severity, and distorted vowels) demonstrating 68% or greater exact agreement. Many of the sentence features had exact agreement between at least two raters at levels of 90% or higher.

Table 1.

The percentage of application of Rule 1 (Rule 2 is the complement) and the mean and standard deviation of the absolute difference of original scores from the rule-based score for each feature in the vowel and sentence rating tasks.

Feature	Vowel Rule 1 applied (%)	Vowel M (SD) of absolute difference from rule-based score	Sentences Rule 1 applied (%)	Sentences M (SD) of absolute difference from rule-based score
Roughness	80	0.70 (0.74)	68	0.84 (0.77)
Breathiness	91	0.82 (0.85)	81	0.62 (0.79)
Strain	86	0.51 (0.66)	98	0.57 (0.90)
Pitch variability	79	0.78 (0.82)	98	0.28 (0.60)
Pitch breaks	98	0.67 (0.32)	100	0.30 (0.20)
Pitch level	NA	0.66 (0.71)	82	0.53 (0.70)
Loudness variability	97	0.51 (0.63)	98	0.33 (0.67)
Dysphonia severity	83	0.65 (0.67)	73	0.71 (0.72)
Atypical resonance	73	0.64 (0.65)	77	0.58 (0.63)
Imprecise consonants	NA	NA	90	0.45 (0.62)
Distorted vowels	NA	NA	72	0.64 (0.70)
Irregular articulatory breakdown	NA	NA	98	0.21 (0.61)
Speaking rate	NA	NA	96	0.55 (0.66)
Disturbance of speech rhythm	NA	NA	77	0.60 (0.65)
Atypical intonation	NA	NA	84	0.62 (0.67)
Disfluency	NA	NA	98	0.17 (0.42)
Reduced intelligibility	NA	NA	94	0.54 (0.68)
Overall atypical quality	NA	NA	89	0.52 (0.62)

Note. NA = not applicable.

Once this rule-based score was defined (whether by Condition 1 or 2), we in turn defined a mean absolute difference between each rater's score and the rule-based score. The absolute differences provide another index characterizing the magnitude of disagreement. Across features, we observe mean values for the mean absolute difference that is always less than 1 scale point.

To further help neutralize the effects of rating error, we next explored defining composites of the rule-based feature scores, as opposed to individual rule-based feature scores. To this end, using the rule-based scores, we considered two primary methods of data reduction: (a) latent profile analysis and (b) principal component analysis (PCA). Both procedures provide a smaller number of dimensions by which to understand speaker differences; latent profile analysis emphasizes qualitative discrete speaker types, while PCA characterizes continuous dimensions. As both approaches rendered comparable interpretations, we report on the PCA, given the likely meaningfulness of real continuity in these dimensions across speakers with DS and the anticipated relevance in the continuity in understanding relationships to speech intelligibility (communicative effectiveness feature of reduced intelligibility; see the Appendix). Briefly, PCA seeks to summarize speech features using a smaller number of orthogonal variables. The principal components are defined as a weighted linear combination of the original feature variables, that is,

$$Z_i={\sum }_i{\mathrm{\varphi }}_{\mathit{ik}}{X}_k\mathrm{,}$$ (1)

where Z_i is the ith principal component, X_k s are the original feature variables, and ϕ _ik is the ith principal component weight assigned to feature k. The principal components are sequentially ordered in terms of importance, that is, the first principal component has the largest variance, the second principal component has the second largest variance, and so forth.

Separate analyses were performed using the rule-based scores for features collected from the vowel- and sentence-based analyses. In each instance, the data analyzed were based on the rules-based scores of the features across raters. The features chosen for analysis were those that (a) were independently rated/evaluated as unique features that were present and (b) entailed a rating scale metric that could be unambiguously interpreted with in relation to intelligibility (e.g., high scores reflect a less intelligible utterance). For the vowel-based analyses, a total of eight of the 13 rated features (roughness, breathiness, strain, pitch variability, pitch break, loudness variability, dysphonia severity, and atypical resonance) were entered into the PCA, while for sentences, 13 of the 24 features (roughness, breathiness, strain, pitch variability, loudness variability, dysphonia severity, imprecise consonant articulation, distorted vowels, irregular articulatory breakdown, atypical resonance, disturbance of speech rhythm, atypical intonation, and disfluency) were used as input variables. For each set of features, we extracted components from the interfeature correlation matrix and attended to the principal component eigenvalues in deciding how many components to extract. The resulting component scores from each analysis were in turn correlated with speaker age, sex, and the ratings of features, not entered in the PCA (including pitch level for vowels and the overall communicative effectiveness features of reduced intelligibility and atypical overall quality of speech, as well the rating of speech rate, for sentences).

Results

Using the rule-based strategy scores, we report results for the perceptual ratings of individual features and subsequent principal components with respect to (a) the distributions of individual features and comparisons in terms of severity, (b) the interpretation of components for vowel and sentence features, and (c) estimated intercorrelations between the component scores and other external variables.

Severely Affected Features in the Speech Production Subsystems

The rule-based ratings for the most severely affected features are shown in Table 2 for vowels and Table 3 for sentences. The features are listed in order of the mean level of severity. Features not included in these tables were those that had predominantly ratings of “1” (not present) and, therefore, do not reveal the nature of the speech disorder. The results for vowels (see Table 2) indicate that the majority of speakers were judged to have some degree (i.e., a severity rating other than “1”) of dysphonia and atypical resonance. The results in Table 3 show that for the sentence task, a large majority of speakers with DS had atypical overall quality of speech, imprecise consonants, reduced intelligibility, atypical resonance, distorted vowels, disturbance of speech rhythm, dysphonia, and atypical intonation. These features reflect involvement of all subsystems of speech production.

Table 2.

Shown for the features judged to be most severely affected in the vowel task are the mean ratings and the percentage of ratings for each scale value.

Feature	Mean rating	Rating
		1 Not Present	2 Mild	3 Moderate	4 Severe	5 Very severe
Dysphonia	2.76	3	36	41	19	0
Roughness	2.36	10	48	35	8	0
Breathiness	2.15	35	36	13	16	0
Atypical resonance	2.08	23	47	28	2	0
Pitch variability	1.99	38	29	29	5	0
Loudness variability	1.56	52	40	8	0	0
Strain	1.5	40	17	6	1	0

Table 3.

Shown for the 10 features judged to be most severely affected in the sentence task are the mean ratings and the percentage of ratings for each scale value.

Feature	Mean rating	Rating
		1 Not present	2 Mild	3 Moderate	4 Severe	5 Very severe
Atypical overall quality	3.36	0	25	34	28	13
Imprecise consonants	3.18	2	37	20	22	19
Reduced intelligibility	3.15	8	34	14	24	20
Atypical resonance	2.73	1	44	37	15	3
Distorted vowels	2.66	17	36	22	17	9
Disturbance of rhythm	2.62	10	30	47	13	0
Dysphonia	2.44	8	53	29	10	0
Atypical intonation	2.39	13	46	32	10	0
Roughness	2.22	20	39	38	2	0
Breathiness	1.82	58	13	20	8	0

For the features rated on a bipolar scale, the results are summarized in Table 4 for the vowel and sentence tasks. Briefly, pitch level for the vowel task was judged to be typical for age and sex (a rating of 3) in 40%, lower than typical (a rating of 1 or 2) in 42%, and higher than typical (a rating of 4 or 5) in 18% of speakers. Pitch level for the sentence task was judged to be typical for age and sex in 33%, lower than typical in 51%, and higher than typical in 15%. Speaking rate was judged to be normal (a rating of 3) in 31% of the speakers, somewhat slow (a rating of 2) in 44%, and somewhat rapid (a rating of 4) in 26%. Therefore, the individuals with DS were not uniform in perceived pitch level or speaking rate for sentences.

Table 4.

Percentages of speakers assigned each scale value in the binary scales of pitch level for the vowel and sentence tasks and speaking rate for the sentence task.

Feature	Rating
	1 Very low	2 Low	3 Normal	4 High	5 Very high
Pitch level–vowel	8%	34%	40%	18%	1%
Pitch level–sentences	2%	49%	33%	14%	1%
Speaking rate–sentences	1%	44%	31%	26%	0%

Principal Components Analysis: Vowels and Sentences

Table 5 displays the principal component eigenvalues for the principal components in the vowel and sentences analyses. Adopting an eigenvalue-greater-than-1 criterion for the retention of a component, we observe four-component solutions for both the vowel and sentence analyses.

Table 5.

Eigenvalues from principal components analyses applied to correlation matrix, for vowel and sentences features.

Component	Vowel analysis			Sentence analysis
	Eigenvalue	% Variance	Cumulative %	Eigenvalue	% Variance	Cumulative %
1	2.418	30.225	30.225	3.986	30.664	30.664
2	1.521	19.019	49.244	1.661	12.780	43.443
3	1.203	15.032	64.276	1.389	10.687	54.131
4	1.016	12.703	76.980	1.264	9.726	63.856
5	.783	9.789	86.769	.945	7.267	71.123
6	.470	5.879	92.648	.842	6.479	77.602
7	.373	4.661	97.309	.726	5.588	83.190
8	.215	2.691	100.00	.644	4.953	88.143
9				.560	4.311	92.454
10				.433	3.332	95.787
11				.268	2.064	97.850
12				.182	1.403	99.253
13				.097	.747	100.00

For vowel analysis (see Table 6), the four component loadings of the eight features entered into the PCA reflect the correlations between the features and components; they also define the proportional weight assigned to each feature in defining the component. Principal Component 1 (PC1) was determined to reflect the severity of speech impairment, as it distinguishes speakers high on roughness (.719), dysphonia severity (.702), strain (.655), pitch variability (.650), and loudness variability (.603) from speakers low on these same variables. In contrast, Principal Component 2 (PC2) distinguishes speakers high on breathiness (.783) and atypical resonance (.523) and low on strain (−.594) from speakers low on breathiness and atypical resonance and high on strain. Principal Component 3 (PC3) distinguishes speakers high on pitch break (.637), low on loudness variability (−.505), and high on dysphonia (.446) from speakers low on pitch breaks, high on loudness variability, and low on dysphonia, while Principal Component 4 (PC4) distinguishes speakers high on pitch break (.560) and low on atypical resonance (−.416) from speakers low on pitch breaks and high on atypical resonance. PC1 is interpreted to identify speakers with vocal hyperfunction (roughness, strain, pitch and loudness variability, and dysphonia), while PC2 is interpreted to identify speakers with vocal hypofunction (breathiness but not strain) and with atypical resonance. PC3 relates to aspects of phonatory function and identifies speakers who are dysphonic with pitch breaks and poor control of loudness. PC4, based on pitch breaks and atypical resonance, is not easily interpreted but indicates that features of speech can be differentially affected in DS. The identification of either hyperfunction or hypofunction in phonation is consistent with previous studies reporting one or the other of these characteristics in speakers with DS. (Moran & Gilbert, 1982; Pebbili et al., 2019; Wold & Montague, 1979).

Table 6.

Principal component loading matrix, first four principal components, vowel features.

Feature	Component
	1	2	3	4
Roughness	.719	−.188	.328	−.336
Breathiness	.356	.783	.127	.337
Strain	.655	−.594	.150	−.129
Pitch variability	.650	−.278	−.363	.335
Pitch breaks	−.153	−.114	.637	.560
Loudness variability	.603	.037	−.505	.402
Dysphonia	.702	.394	.446	−.114
Atypical resonance	.210	.523	−.256	−.416

Note. The largest loadings (in absolute magnitude) on each component define importance. Component loadings greater than .4 in absolute magnitude are shown in bold.

Sentence analysis similarly entailed evaluating the relative contributions of features in defining each component (see Table 7). The interpretation of the component loadings, reflective of the correlations between the features and components, is likewise indicative of general speaker characteristics distinguished by different subsets of features. PC1 distinguishes speakers high on imprecise consonant articulation (.865), distorted vowels (.845), disturbance of speech rhythm (.777), atypical intonation (.694), breathiness (.650), atypical resonance (.648), and dysphonia severity (.635) from speakers low on these same variables. PC2 distinguishes speakers high on roughness (.828), dysphonia severity (.551), and disfluency (.628) from speakers low on these same variables. PC3 distinguishes speakers high on irregular articulatory breakdown (.716) and strain (.680) from speakers low on these same variables, while for PC4, speakers high on loudness variability (.714) and pitch variability (.705) are distinguished from speakers low on these same variables.

Table 7.

Principal component loading matrix, first four principal components, sentence features.

Feature/Component	1	2	3	4
Roughness	.194	.828	−.022	−.165
Breathiness	.650	−.060	−.353	−.116
Strain	.052	.250	.680	.006
Pitch variability	.343	−.031	−.100	.705
Loudness variability	−.026	.025	.214	.714
Dysphonia	.635	.551	−.287	−.133
Imprecise consonants	.865	−.309	.053	−.077
Distorted vowels	.845	−.219	.140	−.055
Irregular articulatory breakdown	.172	.228	.716	−.193
Atypical resonance	.648	−.097	−.147	−.123
Disturbance of rhythm	.777	−.018	.180	.099
Atypical intonation	.694	.080	.104	.195
Disfluency	−.061	.628	−.253	.300

Note. The largest loadings (in absolute magnitude) on each component define importance. Component loadings greater than .4 in absolute magnitude are shown in bold.

These principal components are not easily interpreted, but it is noteworthy that, as in the case of the vowel analysis, both phonatory hyperfunction and hypofunction emerge in the first three components. PC1 is associated with breathiness (interpreted to be hypofunction), whereas PC2 is associated with roughness (interpreted to be hyperfunction) and PC3 is associated with strain (also interpreted to be hyperfunction). It is noteworthy that the phonatory features associated with PC1 are different between the vowel versus sentence tasks, with roughness for vowels and breathiness for sentences. This result may indicate differences in vocal function related to task requirements. More generally, PC1 may reflect an overall greater severity of disorder across features and subsystems (phonatory, articulatory, resonatory, and suprasegmental). That is, this component is related to a greater involvement of all speech subsystems, rather than a narrow set of subsystem difficulties.

Linking Principal Components to Other Variables

We examined intercorrelations between each of the principal component scores and other speaker variables, as well as those features with bipolar ratings that were not included in the PCA. In the vowel analysis, these variables included the sex and age (in months) of the speaker, as well as the pitch level bipolar rules score and the deviation of the pitch level rules score from 3 (the value for normal or typical pitch). We observed no significant correlations between sex and any of the components, while age was significantly correlated only with PC4 (high on pitch break and low on atypical resonance; r = −.40, p = .001), implying younger speakers tended to achieve higher scores on this component (implying greater severity). For the pitch level rule-based score, we observed significant negative correlation with PC2 (r = −.47, p < .001; see Table 6), implying higher breathiness and atypical resonance combined with low strain is associated with lower pitch level. Lastly, for the pitch level absolute deviation from 3, we observed significant correlations for PC2 (r = .30, p = .017) and PC4 (r = −.27, p = .034). Such correlations imply higher breathiness and atypical resonance combined with low strain are associated with greater departures from the normal or typical pitch level, while lower breathiness and atypical resonance combined with high strain are associated with being closer to normal pitch level. In addition, high pitch break combined with low atypical resonance is associated with being close to typical pitch level, while low pitch break combined with high atypical resonance implies being farther away from normal pitch level.

Similar correlations were examined between component scores based on the sentences analysis and the variables of speaker sex, age (in months), reduced intelligibility, and both atypical overall quality of speech rule-based score, the bipolar speaking rate rule-based score and the deviation of the speaking rate rule-based score from 3. For sex, a significant correlation was observed with PC1 (r = .31, p = .006), indicating males had higher component scores (and thus greater severity; see Table 7). For age, we observed a positive correlation with PC2 (r = .25, p = .024) indicating that older individuals show a higher combination of roughness, dysphonia severity, and disfluency. Regarding features related to overall communicative effectiveness with bipolar ratings, both reduced intelligibility and atypical overall quality of speech rule-based scores showed significant correlation with PC1 (r = .85, p < .001, and r = .86, p < .001, respectively), implying higher severity of imprecise consonants, distorted vowels, disturbance of rhythm, atypical intonation, and breathiness are related to decreased overall communicative effectiveness. Reduced intelligibility was also significantly correlated to PC2 (r = −.26, p = .021), as was speaking rate (r = −.29, p = .010), implying lower combinations of roughness, dysphonia severity, and disfluency are related to lower ratings of reduced intelligibility and faster speaking rate.

Finally, in attempting to correlate component scores across both sets of analyses, we saw virtually no significant correlations between components defined from the vowel and sentence features. The only significant correlation observed occurred for PC3 of the vowel analysis with PC2 from the sentence analysis (r = .27, p = .035). While our sample size is not large (and thus power in detecting correlations is low), it would appear that the primary distinguishing characteristics of the feature scores for the vowel- and sentence-based features are largely different when evaluating speech segments. At the same time, it should be kept in mind that the reliability of ratings is also not high, which may also contribute to weaker estimated associations.

Discussion

The major conclusions of this study are that (a) speech disturbances were common, with many perceptual features for both the vowel and sentence tasks judged as present (i.e., atypical) with a severity ranging from mild to very severe, (b) the speech disorder in DS is not homogeneous across individuals with DS, and (c) the speech disorder reflects impairments that are distributed across the subsystems of speech production.

The first result, that speech disorder is common and often severe in the individuals studied in this report, is consistent with several other studies and reinforces the conclusion that speech production is significantly impaired in many individuals, both children and adults, with DS. In this study, atypical overall quality was judged to be present in all of the speakers, and approximately 90% were judged to have some degree of reduced intelligibility for the sentence task. For the vowel task, features of abnormal phonation and resonance were judged to be present for the majority of speakers (see Table 2). For the sentence task, features of articulation (imprecise consonants, distorted vowels), phonation (dysphonia), resonance (atypical resonance), and suprasegmentals (disturbance of rhythm, atypical intonation) were judged as present for most speakers (see Table 3).

The second result, of nonhomogeneous speech patterns, is consistent with reports that emphasize individual differences in DS (Karmiloff-Smith et al., 2016; Thomas et al., 2020). Although studies of speech in DS often have been conducted with the goal of defining a typical or syndrome-defining speech pattern, discrepant results have been reported, as noted in the introduction. This study reveals that a likely reason for these discrepancies is that the speech patterns are not in fact uniform across individuals with DS. There may be some features that occur with high frequency in the speech of persons with DS, but these features await precise description in a large sample of individuals as to their nature and origin. A major conclusion of this article is that speech characteristics are heterogeneous and do not necessarily conform to a syndrome-specific pattern. Bearing this caveat in mind, we can offer the following description of the most typical features of the speech disorder in DS: reduced intelligibility, atypical overall quality of speech, errors in the articulation of consonants and vowels, dysphonia associated especially with low pitch and rough voice quality, abnormal resonance, and dysprosody affecting rate, rhythm, and intonation. This constellation of features highlights the complexity of the speech disorder in DS.

The third conclusion, that impairments are distributed across the subsystems, is consistent with a narrative review (Kent & Vorperian, 2013) and recent research reports (Jones et al., 2019; O'Leary et al., 2020; E. M. Wilson et al., 2019a, 2019b) and has substantial implications for clinical assessment and treatment. Specifically, it is important to recognize the manifold nature of the speech disorder and to consider the profile of disturbances and limitations across the speech subsystems in individuals with DS. This conclusion indicates the need to address subsystems of speech production in both clinical assessment and research studies. Speech subsystem impairments are discussed in more detail in the following section, which considers issues involved in the measurement and interpretation of subsystem performance.

Speech Subsystem Analysis

Considered in subsystems, the severity of the perceptually rated speech dimensions for the sentence task was (from highest to lowest) articulation (imprecise consonants, distorted vowels), suprasegmentals (disturbances of rhythm, atypical intonation), resonance (atypical resonance), and phonation (dysphonia, roughness, breathiness). (The vowel task pertains only to the subsystems of phonation and resonance). Each of these subsystems is discussed below.

Articulatory Subsystem

The features of imprecise consonants and distorted vowels appeared frequently and were among the most severely affected aspects of speech. The imprecise consonants feature was observed in the great majority of the participants, closely in line with the percentages for reduced intelligibility and atypical overall quality. The feature of distorted vowels was somewhat lower in severity but was still noted in the majority of speakers. The nature of articulatory errors has been studied for the most part with traditional tests of articulation (Kumin et al., 1994; Roberts et al., 2005; Sommers et al., 1988) that have addressed questions such as whether the speech sound errors are better characterized as delayed or atypical in nature.

Acoustic and physiological studies have the potential to provide additional supplemental information. Carl et al. (2020) reported that low vowels are more centralized and variable in speakers with DS than in typically developing controls. This result parallels the conclusion by Wild et al. (2018) that the low vowels are particularly susceptible to confusion in the speech of persons with DS. Carl et al. also determined that the acoustic data were associated with overall intelligibility scores and that vowel formant dispersion is a sensitive measure to distinguish formant data between speakers with DS and speakers with typical development. Studies using electropalatography have revealed atypical articulatory patterns of consonant production in individuals with DS and have shown that this method can be used in visual biofeedback therapy (Wood et al., 2019). It also has been shown that children with DS may have reduced tongue and lip strength compared to typically developing children (Zarzo-Benlloch et al., 2017). Acoustic and physiological studies may hold the key to distinguishing anatomic versus neurological factors in accounting for articulatory errors (Kent & Vorperian, 2013, 2021).

Suprasegmental Subsystem

As noted in the introduction, aspects of prosody are disturbed in DS, but reports do not agree on the origin or nature of the disturbance. Speaking rate in conversation has been described as rapid (Fawcett & Peralego, 2009; Kumin, 2006) or slow (R. S. Chapman et al., 1998). It may be useful in future studies to distinguish pause time from actual articulation time, given that these may be differentially affected in DS (Jarrold et al., 2004). Although diadochokinetic rate generally has been described as decreased (Brown-Sweeny & Smith, 1997; Hamilton, 1993; Rosin et al., 1988; Rupela & Manjula, 2010; Zarzo-Benlloch et al., 2017), McCann and Wrench (2007) reported a rate similar to that in typically developing children (McCann & Wrench, 2007). The data in this study (see Table 4) indicate that speaking rate varied across individuals, which is further evidence that speech production is heterogeneous in DS.

The suprasegmental (prosodic) features of disturbances of speech rhythm and atypical intonation were judged to be affected to some degree in about 90% of the speakers. This result agrees with previous studies (Stojanovik, 2011; Stojanovik et al., 2016; Zampini et al., 2016; Zanchi et al., 2020), but there is not yet a clear picture of the nature and origin of the prosodic disturbance. Because prosody in a given task can be influenced by linguistic as well as perceptual, cognitive, and physiological (e.g., respiratory or laryngeal) factors, systematic research is needed to understand the basis of the dysprosody. This is particularly important given that childhood apraxia of speech may be part of the speech disorder in DS (Coêlho et al., 2020; Kumin, 2006; Rupela et al., 2016; E. M. Wilson et al., 2019a, 2019b). According to the American Speech-Language-Hearing Association (2007), one of the three primary diagnostic characteristics of childhood apraxia of speech is inappropriate prosody, especially in the realization of lexical or phrasal stress. The ratings performed in this study may not have been sensitive to the presence of an apraxia separate from co-occurring speech disorders such as dysarthria or structurally based disorder.

Resonatory Subsystem

Atypical oral–nasal resonance is a frequently noted feature of speech in DS, but the exact nature and origin of this feature are unclear. In this study, raters often identified the feature of atypical resonance for the sentence task but differed for both the vowel /ɑ/ and sentence task on whether this feature represented hyponasality, hypernasality, or some other type of abnormal resonance (e.g., nasopharyngeal resonance, cul-de-sac resonance). It was for this reason why our data analysis was restricted to the general category of atypical resonance rather than its subtypes. Poor interrater agreement of resonance features has been reported in previous studies of DS (Jones et al., 2019; Montague & Hollien, 1973). The reasons for the lack of agreement may reside partly in the complex nature of oral–nasal resonance as a general phenomenon across speech disorders and in the various factors that influence oral–nasal coupling in DS. Concerning the latter, adenotonsillar hypertrophy, which is frequently observed in children with DS, is likely to result in hyponasality because of obstruction of the velopharyngeal port. Surgery in the form of adenoidectomy and tonsillectomy may result in the opposite feature of hypernasality as a transient or more longstanding problem. In addition, abnormal development of the paranasal sinuses and nonpneumatization of one or more sinuses (J. D. R. Miller et al., 1986; Shott, 2006) could result in altered patterns of resonance in DS. The acoustic effects of the paranasal sinuses have been described for typically developing speakers (Dang et al., 1994; Havel et al., 2016; Lindqvist-Gauffin & Sundberg, 1976). Further study may reveal if at least part of the atypical resonance in DS is attributable to maldevelopment and nonpneumatization of the paranasal sinuses.

The classification of oral–nasal resonance disorders derives largely from studies of cleft palate (Kummer, 2011), and it is not certain that such classification applies to DS, which may involve unique anatomic and neurological features affecting oral–nasal coupling. The atypical oral–nasal resonance in DS is sometimes identified as nasopharyngeal resonance or cul-de-sac resonance. These are described impressionistically (see Kummer, 2011, for details). Cul-de-sac resonance presumably results when sound resonates in a cavity but is not transmitted beyond the cavity because of obstruction at the cavity's exit point. Speech is described as muffled and low in volume. It has been called “potato-in-the-mouth” speech. The Prosody-Voice Screening Profile (Shriberg et al., 1992) defines nasopharyngeal resonance as a “muffled,” “back of the throat” quality consistent with the percept of “sluggish or imprecise tongue movement” that may characterize the speech of persons with DS. However, the term nasopharyngeal resonance is not widely used in describing speech disorders except in relation to the profile developed by Shriberg et al. (1992).

Acoustic studies may shed light on this issue, but few relevant studies have been reported. Fourakis et al. (2010) concluded that the feature of nasopharyngeal resonance in DS is associated with lowering of the second formant frequency, which they interpreted to be the result of backing of the tongue. However, the lowering of the second formant could also be regarded as a feature of vowel articulation and not atypical resonance per se. Further study of the acoustic correlates should consider poles and zeroes associated with a shunt resonator. In a study of nasalance, Qing and Jiao (2012) observed that, before the age of puberty, children with DS had a significantly lower nasalance value on nonnasals and a significantly higher value on nasals. The pattern changed after the age of puberty, and the authors concluded that nasalance depended on the age and health of the children with DS. This observation supports the need for developmental studies of speech production in DS.

Respiratory/Phonatory Subsystem

The presence of a voice disorder in DS is well attested in the literature (Albertini et al., 2010; Kent & Vorperian, 2013) and often is said to be distinctive of the syndrome. Moran (1986) reported that speech-language clinicians can distinguish individuals with DS from individuals without DS (but with dysphonia) by listening only to sustained vowels. However, as indicated in the introduction, published reports disagree on the nature of the phonatory disorder. Even when listeners agree that dysphonia is present, they do not necessarily agree on the dimensions or features of the voice disorder. Pentz and Gilbert (1983) reported that when listeners rated voice samples from children with DS using the Wilson Voice Profile (F. Wilson, 1971), the only subscale that distinguished the children with DS from controls was that of severity. In the current study, we interpret the PCA of the vowel data to mean that some speakers have vocal hyperfunction while others have vocal hypofunction. The voice features in DS vary across individuals, across speaking tasks, and possibly even within individuals, and further study is needed to determine how these variations relate to age, sex, and possibly other factors. The origin of these different vocal patterns could lie in anatomic features such as laryngomalacia (Bertrand et al., 2003) or the direct influence of hypotonia or, more indirectly, to efforts to compensate for it. Pryce (1994) concluded that individuals with DS use higher energy levels of muscular activation for phonation compared with neurotypical controls, perhaps in an effort to overcome hypotonia.

Vocal pitch or its acoustic correlate, vocal fundamental frequency, has been reported to be higher in speakers with DS in several studies (Albertini et al., 2010; Corrales-Astorgano et al., 2018; M. T. Lee et al., 2009; Rochet-Capellan & Dohen 2015), but lower in others (Moura et al., 2008). For vowels in this study, pitch level was judged to be typical in 40%, higher than typical in 17%, and lower than typical in 43%, while for sentences, pitch level was judged to be typical in 33%, higher than typical in 15%, and lower than typical in 51% (see Table 4). For both the vowel and sentence tasks, then a lower than typical pitch predominated but a higher than typical pitch also occurred. The explanation of atypical pitch levels could be in laryngeal dysmorphology, motor disorder (e.g., laryngeal hypotonia), respiratory dysfunction, or any combination of these. As mentioned for other features, it should be recognized that pitch level can be atypical for age and sex in either direction (high or low).

In view of these discrepancies in perceptual descriptions of voice in DS, other methods of study, such as acoustic and aerodynamic measures, may be more informative. Unfortunately, the instrumental studies reported to date also show lack of agreement. No single set of acoustic variables has emerged to define voice in DS (Albertini et al., 2010; Moran, 1986; Moran & Gilbert, 1982), but Pentz and Gilbert (1983) concluded that children with DS, compared to healthy controls, had increased frequency and amplitude perturbations and higher noise-to-harmonic spectral ratios. In a review of acoustic and physiological studies, Krishnamurthy and Ramani (2020) concluded that there were no significant differences in acoustic variables/characteristics of voice between children with DS and typically developing children but that aerodynamic data indicated a significant difference relating to “glottal valving.” In a recent study that used acoustic measures to infer laryngeal biomechanics in a group of eight 5- to 6-year-old children with DS, the main vocal features identified were deviations in the fundamental frequency, vocal fold tension, and phonation tremor, which were interpreted to result from neuromotor dysfunction (Hidalgo-De la Guía et al., 2021). The path to progress may rely on a more extensive set of acoustic and aerodynamic measures obtained from a larger participant sample, with the recognition that persons with DS may manifest either laryngeal hyperfunction or hypofunction.

Distinguishing Dysmorphology and Disordered Motor Control

As noted in the introduction, the intelligibility and quality of speech in DS are likely affected by both dysmorphologies and aberrant motor control. The dysmorphologies are numerous and variable across individuals. Some of the most frequently noted of these are relative macroglossia (also called pseudomacroglossia or false macroglossia), shortened midface skeleton, short and narrow palate, mandibular dysplasia, dental malocclusion, and laryngomalacia (Kaczorowska et al., 2019; Sforza et al., 2012; Uong et al., 2001). In combination, these anatomic deviations could contribute to disorders of phonation, resonance, and articulation, all of which are common in DS (Jones et al., 2019, and this study). A better understanding of anatomic determinants of speech problems may come from studying the effect of interventions, such as rapid maxillary expansion (Moura et al., 2005), orthodontic palate plate therapy (Bäckman et al., 2007; Carlstedt et al., 2003; Javed et al., 2018), visual biofeedback therapy based on electropalatography (Wood et al., 2019), and intensive voice treatment (Langlois et al., 2020; Mahler & Jones, 2012). These interventions have shown promise in improving certain aspects of oral function in at least some children with DS. It is particularly important that future studies of these therapies provide precise descriptions of changes in speech articulation, preferably by instrumental (acoustic or physiological) and auditory-perceptual methods. It would be particularly helpful to correlate imaging data on vocal tract dimensions with acoustic measures such as vowel formant frequencies. Real-time magnetic resonance imaging correlated with acoustic measures would be helpful in this regard.

Also, frequently noted in DS are neurological or motor control abnormalities, including hypotonia, dysarthria, apraxia of speech, and other motor disorders, which may co-occur. Hypotonia or reduced muscle tone is a common clinical sign and is frequently mentioned as a factor in accounting for motor dysfunction in DS (Santoro et al., 2020). However, as Latash et al. (2008) commented, although hypotonia may be a commonly used clinical term, it does not have (a) a clear definition, (b) transparent interpretation as to underlying mechanisms, and (c) a universally accepted method of measurement. With respect to the orofacial system, Chu and Barlow (2016) remarked that “current experimental data which indicate a causal relation between hypotonia and speech impairment are generally lacking for individuals with DS” (p. 26). Hypotonia cannot be discounted as a factor in explaining speech motor problems in DS, nor should it be accepted as the primary factor until confirmatory data are obtained from sophisticated biomechanical assessments.

The speech disorder in DS has been associated with a dysarthria and apraxia of speech. The type of dysarthria has seldom been identified in research reports, but recent studies point to ataxic (E. M. Wilson et al., 2019b) or flaccid (Langlois et al., 2020). According to a technical report from the American Speech-Language-Hearing Association (2007), childhood apraxia of speech is characterized by some combination of the following features: errors in the production of vowels and consonants, unusual and variable errors in repeated attempts at pronouncing syllables, abnormalities in prosody, and poor transitions between sounds and syllables when coarticulation is required. Other features associated with CAS include a limited phonemic repertoire and reduced articulatory precision with increases in word length and speaking rate (Teverovsky et al., 2009). However, precise description of the motor disorder is difficult because dysarthric and apraxic features may co-occur and potentially interact. Rupela et al. (2016, p. 9) summarized the issue: “Due to overlapping symptoms, it remains difficult to discern which of the two disorders contributes more to the motor speech difficulty observed in DS.” It is also possible that the dysarthric and apraxic features interact to produce a motor disability that is specific to DS. Given the current state of knowledge, there is little basis to distinguish co-occurring motor disorders in DS or to separate their effects from concomitant dysmorphologies. Determining mutual exclusivity of features for different etiologies may be unlikely from perceptual data alone. Furthermore, in order to assess global speech patterns, it is necessary to have a more complete picture that accounts for language development, including vocabulary, syntax, and pragmatics. For example, difficulties with syntax is a particular linguistic challenge for individuals with DS that could influence speech patterns related to fluency and prosody (Andreou & Chartomatsidou, 2020).

Although studies of speech in DS usually focus on impairments, it is important to note that some children and adults with DS are highly intelligible and have essentially normal perceptual attributes of speech. In this study, 10% of the speakers had highly intelligible speech. Much can be learned from the careful study of these individuals to determine the anatomic, motor, and other correlates of their speech abilities. Phenotypic variation is at once a challenge and an opportunity in understanding the speech disorder in DS. Efforts to establish patient registries, research databases, and biobanks are expected to lead to improvements in basic science and clinical intervention (Oster-Granite et al., 2011).

Limitations of Auditory-Perceptual Judgments

This study is further evidence of the difficulty that listeners may have in rating the multiple features of a complex speech disorder. Although the raters in this study were highly experienced in making auditory judgments of various speech disorders and were familiar with the speech patterns of individuals with DS and implemented consensus training prior to their ratings, their intra- and interrater reliability was poor for many of the rated features for both the vowel and sentence tasks (with somewhat better reliability for the latter, perhaps because of the greater length of the stimulus). This was not unexpected, as variability in auditory-perceptual judgments of disordered speech is a longstanding problem that becomes especially severe when multiple features are affected (Casilio et al., 2019; Jones et al., 2019; Kent, 1996; A. Lee et al., 2020). Efforts to reduce variability include training of raters and use of perceptual standards to stabilize judgments. These have varying levels of success. Perceptual standards have limited value when multiple features can depart from normal or typical. Statistical methods used to deal with interrater variability include use of the mean (Casilio et al., 2019) or outlier (“odd man out”) procedures such as calculating standardized (z) scores and then excluding values that are 3.29 SDs above or below the mean (Casilio et al., 2019). Neither of these two methods as usually implemented is suitable for this study because our data pertain to only three raters (making calculation of means and standard deviations of dubious value). Similarly, low kappa values have been reported in other studies using auditory-perceptual ratings of speech and voice. For example, in a study of auditory judgments of dysphonia by experienced clinicians using three different voice rating scales, Webb et al. (2004) reported kappa values that were often less than .2 (kappa values are not directly comparable across studies, but the point remains that low values are common in perceptual ratings).

Casilio et al. (2019), in a study of auditory-perceptual features of connected speech in aphasia, commented that, “Interrater reliability was good or excellent for most features for both researchers and student clinicians, so long as scores were averaged across multiple raters. Not surprisingly, scores from single raters were less reliable, even when those raters were experienced researchers” (p. 560). However, as noted earlier, means for a small number of raters can be strongly influenced by an exceptional value (outlier).

Perhaps the greatest source of difficulty is that ratings of any given feature may be influenced by other concurrent atypical features, especially if the other features are severely affected. This problem has been noted in perceptual ratings of voice (Kreiman & Gerratt, 2011; Oates, 2009) and hypernasality–hyponasality (K. L. Chapman et al., 2016; Imatomi, 2005; A. Lee et al., 2020; Tardif et al., 2018). Kreiman and Gerratt (2011) wrote that their findings “are consistent with the view that cognitive and perceptual factors, in addition to random and criterion errors, are in play when listeners judge the quality of complex auditory stimuli like voices” (p. 811). Because auditory-perceptual judgments are one of the main tools of the speech-language clinician, concerns about the validity and reliability of these judgments are of high clinical relevance. This study used rule-based scores as a means of contending with the variability of individual rater scores. In addition, the PCA uncovered patterns among the rated features.

The validity and reliability of perceptual ratings can be affected by several factors, but especially the ability of the scale to represent gradations for the relevant features. For an interval scale, a decision must be made regarding the number of scale values to be used in perceptual ratings. G. A. Miller (1956) concluded that human observers tend toward 7 points (± 2) of gradation in tasks of short-term memory and categorization. As noted by Ludlow et al. (2019), “Although the 7-point scale at one time appeared to be almost a default approach in clinical applications, 3-, 4-, or 5-point scales are increasingly common in specialties such as speech-language pathology, voice disorders, neurology, rehabilitation medicine, and physical therapy” (p. 228). Studies of the optimal number of response categories in rating scales have yielded discrepant results, but Preston and Colman (2000) concluded that reliability increases with number of scale values up to about 7. The choice of 5 scale values in the present research was based largely on current trends in clinical applications and considerations of reliability.

Another consideration is that the features of speech and voice disorder can be either prothetic or metathetic (Stevens, 1946). Prothetic attributes (e.g., sound intensity) are additive or quantitative, whereas metathetic attributes (e.g., pitch) are substitutive or qualitative. Prothetic dimensions are more appropriately scaled with direct magnitude estimation than with equal-appearing interval scaling. However, the latter appears to have overwhelming application across clinical specialties, probably for reasons of convenience and economy. Especially when several features are to be rated, direct magnitude estimation can be difficult to implement (Fukuda et al., 2012; Moskowitz, 1977). Another possibility is to use visual analog scales, such as those incorporated in the Consensus Auditory-Perceptual Evaluation of Voice (Kempster et al., 2009).

Ultimately, the auditory-perceptual description of speech in DS must contend with several limitations or challenges, including (a) inter- and intrarater variability, (b) the multidimensionality of the speech disorder, (c) possible heterogeneity of speech characteristics, and (d) the syndromic features that distinguish DS from other disorders (e.g., the combination of craniofacial dysmorphologies and neurological problems). As mentioned previously with respect to speech subsystems, acoustic and physiological methods may resolve some of the uncertainties and limitations encountered in auditory-perceptual descriptions. A combination of methods has been shown to be informative in the study of dysphonia (Eadie & Doyle, 2005) and is likely to benefit the study of speech in DS. Such an approach has been advocated for the general assessment of speech disorders (Howard & Heselwood, 2011; Kent & Rountrey, 2020; A. Lee et al., 2020).

Implications for Future Studies

It is important to study speech profiles in a large sample of individuals with DS in a life span perspective that recognizes the possibility of heterogeneity among affected individuals. A combination of anatomic, acoustic, and physiological methods is needed for a deeper understanding of the complex speech disorder in this population. Auditory-perceptual methods serve to identify salient characteristics of the speech disorder but cannot in themselves reveal the full details of the complexity of speech production and can only go so far in revealing the factors contributing to the affected speech features. A particular challenge is to determine the roles of dysmorphology, motor speech disorder, and phonological disorder (to name only a few factors) in accounting for the speech difficulties. This information is critical for the development of evidence-based interventions (Ibrahim et al., 2019; Kent & Vorperian, 2021). The heterogeneity warrants a personalized or individualized approach to treatment and prevention, an idea that is consistent with recent reports on intervention in DS (McDaniel & Yoder, 2016; O'Leary et al., 2020) and for precision medicine in general (Gillman & Hammond, 2016). Personalized or precision treatment could be designed in part with consideration of strengths and weaknesses in the subsystems of speech production in each individual. Selection of particular interventions would be conditioned by the speech characteristics and the communication goals of a given person. It is likely that information and communication technologies will play an increasingly important role in education of individuals with DS. In a recent review, it was concluded that the primary educational foci of information and communication technologies in children with DS were reading–writing, communication and language, and motor skills (Fernández Batanero et al., 2019). Enhancing the ease and success of person–machine speech communication would be a decided boon in the further implementation of these technologies.

Appendix

Definitions of Auditory-Perceptual Features Used in Rating Sustained Vowel /ɑ/ and Sentences

Asterisk (*) denotes features used for vowel ratings. All features, except speaking rate and pitch level, were judged on a 5-point equal-appearing interval scale where 1 indicates feature is not present and values 2 through 5 indicate degree of severity. For speaking rate the 5-point scale is bipolar with one pole (1) being too slow and the other (5) too fast. For pitch level, one pole (1) is too low and the other (5) is too high.

1. Features related primarily to phonation and respiration.

 • Roughness*: irregularity in the vibratory pattern of voice, usually related to cycle-to-cycle variations in fundamental period and/or amplitude. Synonyms –creaky, harsh, raspy, glottal fry. Distinguished from voice breaks, which are interruptions of phonation.

 • Breathiness*: friction resulting from escape of air between the vocal folds during phonation. Synonyms -murmured, whispery.

 • Strained*: excessive vocal effort (hyperfunction). Synonym –pressed.

 • Pitch variability*: unusual or unexpected changes in vocal pitch without voice breaks.

 • Pitch break*: interruption of voicing, with voicing resumed after the break.

 • Loudness variability*: unusual or unexpected changes in vocal loudness without voice breaks.

 • Dysphonia severity*: global, integrated impression of voice disorder.

 • Pitch level (bipolar)*: appropriateness of pitch level for age and sex.

2. Features related primarily to supralaryngeal articulation and resonance.

 • Imprecise consonants: consonants are produced inaccurately or with reduced clarity.

 • Distorted vowels: vowels produced inaccurately.

 • Irregular articulatory breakdown: a speech disturbance taking the form of groping articulation or a sudden and inconsistent telescoping of one or more syllables.

 • Atypical resonance*: unusual or unexpected resonance quality that can take the form of hyponasality, hypernasality, cul de sac resonance, or other feature.

 • Hypernasality*: excessive nasal resonance.

 • Hyponasality*: reduced nasal resonance.

 • Cul de sac resonance*: resonance of sound in a cavity (oral, pharyngeal or nasal), as sometimes associated with enlarged tonsils which block the airway.

 • Other atypical resonance*: resonance imbalance other than those above.

3. Features related primarily to suprasegmentals (prosody).

 • Speaking rate (bipolar): the tempo of speech production, usually related to the number of words or syllables produced in a unit of time. Rate deviations can be either too slow or too fast, so that this feature is rated on a bipolar scale.

 • Disturbances of speech rhythm: irregularities in the normal rhythm or stress pattern of speech (e.g., the typical alteration of strong and weak syllables in American English).

 • Atypical intonation: pattern of pitch change (rises and falls) is not appropriate for the utterance (e.g., monotone, excessive pitch change within or across syllables).

4. Features related to overall communicative effectiveness

 • Reduced intelligibility: speech is difficult to understand.

 • Atypical overall quality of speech: speech is unnatural or unusual. Synonym: bizarre.

 • Disfluency: a disruption in the flow of speech, which may take the form of repetitions, hesitations, or prolongations. Disfluencies can disrupt rhythm but not necessarily so.

Funding Statement

Research reported in this publication was supported by awards from National Institutes of Health: Grant Number R01DC006282 and 3R01DC006282-15S1 (Vorperian, PI) from the National Institute on Deafness and Other Communication Disorders, and Waisman Center core grant U54 HD090256 (Chang, PI) from the National Institute of Child Health and Human Development. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.