Traditional and Visual-Acoustic Biofeedback Treatment via Telepractice for Residual Speech Sound Disorders Affecting /ɹ/: Pilot study

Abstract

Purpose:

This study aimed to examine the feasibility of telepractice delivery of a treatment package including visual-acoustic biofeedback and motor-based treatment for residual speech sound disorder affecting /ɹ/ in school-age children. The overall study used a single-case randomization design; however, this preliminary report will simply quantify changes in accuracy before and after completion of the treatment package. The present analysis did not differentiate between the relative contributions of biofeedback and motor-based treatments.

Method:

Seven children aged 9–14 received speech therapy for /ɹ/ distortions via telepractice. The study design consisted of three phases: baseline (four sessions), treatment (20 sessions), and post-treatment (three sessions). Treatment included two sessions weekly for a duration of 10 weeks. The participants received one motor-based/non-biofeedback session and one visual-acoustic biofeedback session per week. The order of treatment within each week was randomly determined prior to the start of therapy. Overall progress was assessed using untrained listeners’ ratings of word probes administered in the baseline and posttreatment phases.

Results:

Findings revealed that six of the seven participants showed a clinically significant response to the overall treatment package, although the magnitude of individual responses varied across speech contexts (consonantal and vocalic) and participants.

Conclusion:

The present results suggest that a treatment combining visual-acoustic biofeedback and motor-based treatment for residual /ɹ/ errors treatment can be effectively delivered via telepractice. Considerations for technology setup and treatment protocols are provided.

Introduction

Telepractice delivery of speech therapy services was increasing even prior to the COVID-19 pandemic. Telepractice can increase flexibility for clients and clinicians (e.g., Pullins & Grogan-Johnson, 2017), making it a valuable complement to in-person service delivery. For individuals in rural areas or for whom transportation is a barrier, telepractice can increase access to certified speech-language pathologists. Moreover, telepractice may enable greater specialization among clinicians (e.g., for cleft lip/palate, Whitehead et al., 2012) because the recruitment of clients is geographically less limited. In the context of COVID-19, telepractice became essential due to social distancing requirements. Although certain interventions, such as ultrasound biofeedback, cannot currently be implemented through telepractice, many other approaches remain available. This paper explores the feasibility of the delivery of visual-acoustic biofeedback and motor-based treatment for older school-aged children and adolescents with speech sound disorder (SSD).

Residual speech sound disorders

Deviations that affect the intelligibility and accuracy of children’s speech can lead to social isolation or bullying (Hitchcock et al., 2015; Crowe Hall, 1991), with potentially life-long consequences for educational and occupational outcomes (McCormack et al., 2009) and mental health (McAllister et al., 2023). Between 2–5% of speakers are estimated to exhibit a residual speech sound disorder (RSSD) that persists through adolescence or adulthood (Shriberg, 2009; Culton, 1986; Flipsen, 2015). RSSD is particularly challenging to treat, as speech-language pathologists frequently discharge clients with unresolved speech deviations (Ruscello, 1995).

Distortions affecting the American English /ɹ/ sound have been characterized as the most common form of RSSD and are widely acknowledged as some of the hardest speech sounds to remediate, often proving resistant to traditional motor-based therapy techniques (Ruscello, 1995; Shuster et al., 1995). Children’s difficulty with /ɹ/ production is commonly attributed to the articulatory complexity of the sound (Gick et al., 2007): while most speech sounds are formed with a single major constriction or narrowing of the vocal tract, /ɹ/ features near-simultaneous anterior and posterior lingual constrictions (e.g., Delattre & Freeman, 1968); /ɹ/ can also be produced with slight labial constriction (King & Ferragne, 2020). In addition, there is a high level of both within- and between-speaker variability in tongue configuration for /ɹ/ (Delattre & Freeman, 1968; Tiede et al., 2004; Zhou et al., 2008). Major variants include retroflex /ɹ/, in which the tongue tip is raised near the alveolar ridge, and bunched /ɹ/, in which the tongue tip is down but the tongue blade and/or body are raised in the vicinity of the palate. However, there is a great deal of variability between these extremes, and some tongue shapes do not fit well into either category (Boyce, 2015). Due to differing coarticulatory pressures, many speakers use distinct tongue shapes across phonetic contexts, such as onset versus postvocalic position (Mielke et al., 2016; Stavness et al., 2012); which may play a role in the observation that children’s age of mastery for /ɹ/ often differs by syllable position (Elbert & McReynolds, 1975). Despite this articulatory heterogeneity, /ɹ/ is characterized by a stable acoustic target: relative to other speech sounds, /ɹ/ is set apart by its low frequency of the third formant, F3. Visual-acoustic biofeedback, which forms the focus of the present study, aims to capitalize on this acoustic stability by providing a visual display of the spectral characteristics of the /ɹ/ sound.

Biofeedback treatment for RSSD affecting /ɹ/

Recent studies have provided evidence that biofeedback-enhanced interventions may help eliminate /ɹ/ distortions in children with RSSD (McAllister & Hitchcock, 2012; McAllister Byun & Campbell, 2016; McAllister Byun et al., 2016; Hitchcock et al., 2017; Preston et al., 2013; Preston et al., 2014). Biofeedback uses instrumentation to create a real-time representation of certain aspects of speech that are typically outside the speaker’s conscious control (Volin, 1998). A model representing the desired production is also displayed, so the learner can be cued to modify their output to match their real-time display to the target image. Various technologies can be used to deliver biofeedback, including ultrasound imaging (e.g., Sugden et al., 2019), electropalatography (e.g., Hitchcock et al., 2017) and electromagnetic articulography (Vick et al., 2017). However, none of these articulatory imaging technologies is amenable to use over telepractice. To explore the possibility of telepractice delivery of biofeedback treatment, we focused on visual-acoustic biofeedback, in which learners view a dynamic image of the acoustic signal of their speech (Hitchcock et al., 2023). Figure 1 shows the real-time linear predictive coding (LPC) spectrum of the speech signal that is displayed by the Sona-Match module of the PENTAX Medical Sona-Speech II software, which is the biofeedback technology used in the present study.

Figure 1:

Real-time linear predictive coding (LPC) spectrum of the speech signal displayed by Sona-Match module of the PENTAX Medical Sona-Speech II software.

Several in-person small scale studies over the past 25 years indicate the effectiveness of visual-acoustic biofeedback (McAllister Byun & Campbell, 2016; McAllister Byun & Hitchcock, 2012; McAllister Byun, 2017; Shuster, et al., 1995). McAllister Byun and Hitchcock (2012) reported that in a study of 11 children who received an initial period of traditional motor-based treatment followed by visual-acoustic biofeedback, significant improvements were identified in perceptual and acoustic measures of /ɹ/ production only after the transition to biofeedback for all but one participant. Additionally, in a different single-case randomization study, three out of seven participants showed a statistically significant advantage for visual-acoustic biofeedback compared to traditional treatment, while none showed a significant advantage in the opposite direction (McAllister Byun, 2017). Lastly, in a single-case experimental study of 11 children with RSSD (McAllister Byun & Campbell, 2016) found a significant interaction between treatment conditions and order. Specifically, participants who received a period of visual-acoustic biofeedback followed by a period of traditional motor-based treatment exhibited significantly larger effect sizes than those who received the same treatments provided in the reverse order.

Telepractice delivery of treatment for speech sound disorders

In a 2017 systematic review investigating speech or language interventions using telepractice for school-aged children with SSD, Wales and colleagues (2017) identified seven studies. The general conclusion was that outcomes appear to be mostly equivalent across telepractice and in-person contexts. For example, similar outcomes were observed in 13 children with SSD from two different schools: one staffed in person and one via telepractice (Grogan-Johnson et al., 2011). In a follow-up randomized controlled study (Grogan-Johnson et al., 2013), 14 children aged six–ten years were randomly assigned to receive speech treatment via telepractice or in-person over five weeks; both groups improved significantly, with no significant between-group difference. Similarly, a retrospective analysis of existing databases (Coufal et al., 2018) found no difference in functional outcomes between children ages six–nine years who had received treatment for SSD via telepractice (n=428) and in person (n=1331). Positive client and parent satisfaction outcomes have also been documented (e.g., Whitehead et al., 2012; Pennington et al., 2019). In contrast, a 2020 systematic review reported that treatment for SSD may yield smaller effect sizes via telepractice versus in-person (Jung et al., 2020). Overall, although a number of studies have supported the viability of telepractice delivery of treatment for SSD, there is a need for further research.

The increased prominence of telepractice treatment during and following the COVID-19 pandemic (Aggarwal et al., 2021; Campbell & Goldstein, 2021) has highlighted the need to expand on the existing evidence base for telepractice treatment for SSD (Wales et al., 2017; Grogan-Johnson et al., 2011; Coufal et al., 2018). Some recent studies have advanced the field’s understanding of remote assessment of SSD but have not addressed the topic of treatment (Bartamay et al., 2023; Campbell & Goldstein, 2022). Most telepractice studies on SSD have either not reported the specific treatment approach utilized (Coufal et al., 2018) or have relied on traditional articulation therapy (Grogan-Johnson et al., 2013). Thus, there is a lack of high-quality experimental evidence regarding the relative or absolute efficacy of different treatment approaches for SSD in the context of telepractice.

Telepractice and visual-acoustic biofeedback

Unlike other forms of biofeedback, visual-acoustic biofeedback (VAB) can, in principle, be delivered via telepractice, as was necessary during social distancing for COVID-19. We are aware of only one previous study (Peterson et al., 2022) that specifically investigated telepractice using VAB treatment for children with residual speech sound disorders (RSSD), a diagnostic category described as particularly challenging to remediate (Ruscello, 1995). Peterson and colleagues (2022) reported that all four participants in their study demonstrated substantial gains in rhotic accuracy after 16 sessions of treatment that were randomly assigned to feature VAB or traditional motor-based treatment. Interestingly, while the average overall effect size was larger than that of a comparable previous study in an in-person setting (McAllister Byun, 2017), fewer participants showed a significant advantage for biofeedback over traditional treatment. The authors speculated that some technological limitations of VAB delivery via telepractice, such as a lag between the participant’s production and the response of the visual display, may have had a diluting effect on outcomes. In light of these findings, further empirical investigation is essential to understand the effects of biofeedback treatment in the context of telepractice. If found to be effective, telepractice delivery of VAB could facilitate access to biofeedback treatment in a larger number of children with RSSD.

In the present study, children with RSSD received both VAB and traditional motor-based treatment in a randomization design similar to that reported by Peterson et al. (2022). However, for this preliminary report, we focused on the perceptual ratings of word-level probes produced in the baseline and maintenance phases, before and after the delivery of all treatments. This allows us to measure the effects of the full treatment package but does not allow us to quantify the relative contributions of the VAB and motor-based treatment components. We intend to pursue this comparison in a companion study using the acoustic analysis of probes administered before and after each treatment session.

Methods

Participants

The study enrolled seven monolingual speakers of American English ranging in age from 9;1 to 14;3 (years; months; mean age 11;0), with an average previous intervention duration of 4 years. (See Table 1 for additional details.) Five participants were male and two were female. Participants were recruited primarily by referrals from local speech-language pathologists or in response to community posts, describing a study focused on the effects of biofeedback intervention delivered via telepractice. Prior to participation, parents/guardians provided informed consent and participants provided assent during a zoom meeting witnessed by a research assistant. The study procedures were approved by the Biomedical Research Alliance of New York (BRANY) along with cooperative agreements from the institutional review boards of Syracuse University, New York University, and Montclair State University.

Table 1.

Participant ID, sex, age in year; months, and previous intervention duration

Participant ID	Sex	Age	Years targeting /ɹ/ prior to study*
3301	M	10;0	2–3 years
3302	F	9;7	1–2 years
3303	F	11;5	6–7 years
3305	M	12;2	4 years
6301	M	10;7	5–6 years
6302	M	14;3	8 years
6303	M	9;1	< 1 year

^*Per parent report

Prior to participation, the parents/caregivers of potential participants completed a questionnaire describing the participant’s speech and health history. Participants with a reported history of hearing loss, neurobehavioral disorder (e.g., Autism Spectrum Disorder, Obsessive-Compulsive Disorder, Tourette Syndrome), developmental disorder (e.g., Down syndrome and cerebral palsy),or hearing concerns were not eligible for inclusion in the present study. Assessment measures were administered using Zoom software. For inclusion, participants were required to pass an oral mechanism exam that assessed labial and lingual range of motion and palatal integrity (based on the absence of hypernasality in non-nasal sentences). Additionally, participants were required to receive a minimum of a scaled score of six on the Recalling Sentences and Formulated Sentences subtests of the Clinical Evaluation of Language Functioning-5 (CELF-5; Wiig et al., 2013), score at or below the 8^th percentile on the Goldman-Fristoe Test of Articulation-3 (GFTA-3; Goldman & Fristoe, 2015), receive a minimum T-score of 37 on the Matrix Reasoning subtest on the Wechsler Abbreviated Scale of Intelligence-Second Edition (WASI-II; Wechsler, 2011), and score below 30% accuracy on a 50-item probe eliciting /ɹ/ at the word level in various phonetic contexts. Participants were screened for signs of Childhood Apraxia of Speech via the Syllable Repetition Task (SRT; Shriberg et al., 2009) and the Inconsistency subtest of the LinguiSystems Articulation Test (LAT; Bowers & Huisingh, 2018). Exclusionary criteria included ≥20% of nonwords produced with speech sound additions on the SRT and ≥ 3 segmental or prosodic errors on the LAT. Non-exclusionary measures include a probe examining rhotic accuracy at the sentence level and a stimulability probe (McAllister et al., 2020). See Table 2 for further details on the assessment.

Table 2.

Assessment results per participant

Participant ID	CELF-5 Recalling Sentences Scaled Score	CELF-5 Recalling Sentences Percentile	CELF-5 Formulating Sentences Scaled Score	CELF-5 Formulating Sentences Percentile	GFTA-3 Standard Score	GFTA-3 Percentile	GFTA-3 Sound Errors	LAT Amount of Inconsistent Productions	SRT Amount of Additions
3301	12	75	11	63	40	<0.1	/ɹ/	0	0
3302	10	50	16	98	40	<0.1	/ɹ/	0	0
3303	8	25	19	99.9	40	<0.1	/ɹ/	0	0
3305	9	37	14	91	40	<0.1	/l/, /θ/, /ɹ/	1	1
6301	14	91	17	99	40	<0.1	/ɹ/	2	1
6302	19	99.9	10	50	40	<0.1	/ɹ/, initial /l/	0	0
6303	11	63	10	50	40	<0.1	/ɹ/	2	2

Abbreviations: CELF-5, Clinical Evaluation of Language Fundamentals/Screener, 5^th Ed.; GFTA-3, Goldman-Fristoe Test of Articulation, 3^rd Ed.; LAT, LinguiSystems Articulation Test; SRT, Syllable Repetition Task

Equipment

Participants were required to have access to a laptop or desktop computer with a Windows or Mac operating system. All sessions were delivered virtually through the Zoom video conferencing platform. A standard headset with a microphone and earphones (Plantronics Blackwire C225) was provided to all participants for use during the study. Probe measures were recorded on the participants’ local devices using Voice Memos (MacOS) or Voice Recorder (Windows) and shared with the study team via upload to a secure cloud content-management website (Box Storage). VAB treatment was delivered using the Sona-Match module of PENTAX Medical Sona-Speech II 3654 v.4 software. Sona-Match software was run on the clinicians’ computers using the Windows 10 operating system. Clinicians shared the Sona-Match screen with the participants via Zoom. Figure 2 shows the clinician’s view of a VAB treatment session via zoom. Sona-Match window settings and /ɹ/ templates were selected for each participant as described in Hitchcock et al. (2023).

Figure 2:

Clinicians’ computer view during treatment sessions using the Zoom video-conferencing platform.

For all treatment sessions, the participants’ and clinicians’ Zoom audio settings were configured to use the “original sound” option to bypass built-in audio settings (e.g., noise suppression and high-pass filtering, which could interfere with the acoustic properties of speech output). A virtual audio mixer (VoiceMeeter) was used to route the participant’s voice to the PENTAX Medical Sona-Speech software operating on the clinician’s computer to generate a visual-acoustic feedback display. Pilot testing indicated that this configuration achieved audio quality sufficient for speech-language pathologists to perform perceptual ratings and provide feedback. While online transmission of audio may introduce a slight temporal lag, our piloting suggested only a minimal degree of lag that was unlikely to disrupt the functional nature of therapeutic exchanges (including in the biofeedback context). To reduce the incidence of temporal lag, the clinicians and participants were required to use an Ethernet cable instead of wireless internet access to maximize bandwidth connectivity (see Appendix A for equipment and audio configurations).

Study Design

The study was conducted online over a 14-week period, including a technology support session, three baseline sessions, an introductory session, 20 treatment sessions (two sessions per week) and three maintenance sessions. Participants were treated individually by one of two certified speech-language pathologists (the first and second authors) who were familiar with VAB treatment. Baseline and maintenance sessions elicited three standard probes featuring 50 words, five sentences, and a stimulability probe (McAllister et al., 2020). All probes contained rhotics in various phonetic contexts; sentence- and word-level probes were presented in random order. Probe items were selected to represent a full range of syllable positions and phonetic contexts because /ɹ/ may be realized with differing accuracy in different environments (Elbert & McReynolds, 1975). Consonantal /ɹ/ was probed in singleton onsets with equal numbers of front- and back-vowel contexts. Vocalic /ɹ/ was probed in stressed and unstressed syllabic contexts (/ɝ/ and /ɚ/), in postvocalic position with front vowels (e.g., /ɪɹ/ and /ɛɹ/) and in postvocalic position with back vowels (e.g., /ɑɹ/ and /ɔɹ/). No feedback was provided during /ɹ/ probe administration. All probes were recorded by the participants and uploaded to a secure Box folder.

Prior to initiating the evaluation sessions, the participants and their parents participated in a technology-support session. Parents were provided with a technology information handout two weeks prior to the technology support session. See Appendix B for a copy of the parent technology handout. During this session, the clinician assessed participants’ computer audio to ensure that it was accurately configured for future sessions. The clinician also guided the participants through the process of recording audio on their computer and uploading the recordings to Box.

Treatment

After the collection of the fourth baseline probe and before the initiation of treatment sessions, participants were provided with two instructional sessions. The first instructional session focused on accurate tongue placement for /ɹ/. Clinicians followed a standard protocol for introducing articulator placement for rhotics (for specific protocol details, see Preston et al., 2020), which was followed by 15 minutes of clinician-directed practice. The second instructional session was designed to familiarize participants with VAB using the Real-Time LPC function on the Sona-Match Module of Sona-Speech II (see Appendix C). In the Sona-Match module, the LPC spectrum has three different viewing/window settings; child, adult female, and adult male. Selection was based upon the age and gender of the individual receiving the intervention. Once the participant began to establish an approximation of a perceptually acceptable rhotic production, the clinician adjusted the participant’s target (e.g., replacing the initial target with one derived from the participant’s best approximation). A participant’s formant pattern target was refined over time as rhotic accuracy improved, providing visual evidence of progression towards age-appropriate production.

Participants were provided with a standard biofeedback training protocol (for additional protocol details and a sample treatment session video, see Hitchcock et al., 2023), which was then followed by 15 minutes of clinician-directed practice using the biofeedback display.

Each participant received one VAB treatment session and one motor-based treatment session per week, individually delivered by one of two certified speech-language pathologists. The study was designed so that VAB and motor-based treatment sessions were as similar as possible (e.g., same target number of trials and schedule of clinician feedback), except that in the VAB condition practice was supplemented with a real-time display of the acoustic signal of speech and compared against a target representing the acoustics of a correct /ɹ/. The order of treatment types within each week was randomized prior to the onset of treatment. Treatment duration was dependent on the participant’s speed of completing treatment trials, but did not exceed 60 minutes in duration. Treatment consisted of an acquisition phase and a generalization phase. Participants remained in the acquisition phase of treatment until cumulative within-session accuracy, based on the treating clinician’s ratings, reached 60%. Once 60% accuracy was achieved, the participant moved to the generalization phase for subsequent sessions of the same treatment type. If the cumulative accuracy for a session then fell below 40%, the participant moved back to the acquisition phase for subsequent sessions of the same treatment type.

In both phases, treatment sessions began with a maximum of 10 minutes of clinician-directed pre-practice in which the participant, with cueing from the clinician, could explore different strategies to produce more accurate /ɹ/ sounds. Participants were provided a model and articulatory placement cues for approximately 75% of the trials during pre-practice. For the biofeedback treatment condition, the VAB display remained available throughout pre-practice. Once participants met a preset performance criterion of three accurate productions per target, they advanced to structured practice; otherwise, clinician-directed pre-practice continued for 10 minutes. In the acquisition phase, pre-practice featured five different syllables, which consisted of /ɝ/ plus one randomly selected syllable from each of four rhotic variants (i.e., prevocalic or postvocalic /ɹ/ with front or back vowels). These syllables were kept constant across sessions during the acquisition phase. In the generalization phase of treatment, two target items per rhotic variant (10 in total) were elicited during pre-practice. In both phases of treatment, clinician-directed pre-practice was followed by the elicitation of 150 trials in structured practice. Structured practice occurred in blocks of 10 consecutive trials of the same target. During the acquisition phase, the same five rhotic syllables targeted in the pre-practice phase were elicited. In the generalization phase, syllables/words targeted in each session were randomly selected from a standard list of words/syllables, and targets could differ between sessions.

In the generalization phase, structured practice featured adaptive difficulty in accordance with principles from the challenge point framework (Guadagnoli & Lee, 2004; Rvachew & Brosseau-Lapré, 2012) using the Challenge Point Program (CPP; McAllister et al., 2021). The CPP software adjusts treatment parameters to make practice more difficult (e.g., reduce the number of trials in which feedback is available) or easier (e.g., restore a higher level of clinician scaffolding) based on the treating clinician’s ratings of accuracy in blocks of treatment. The CPP logs the clinician’s ratings of accuracy and automatically adjusts the difficulty of practice based on the percentage of trials rated perceptually accurate in each block of 10 target stimuli.

Treatment Fidelity

Twenty percent of all treatment sessions were reviewed to evaluate fidelity to the stated treatment protocol. To measure fidelity, video recordings of the structured practice portion of the treatment sessions were rated by the non-treating SLP (i.e., the first author reviewed 12 sessions; the second author reviewed 16 sessions). One session of each type of treatment (i.e., VAB and motor-based treatment) was randomly selected from each half of the study (sessions 1–10 and sessions 11–20), totaling four treatment sessions reviewed per participant (28 sessions total). Fifty trials (one-third of the targets for each assigned session) were reviewed for fidelity. The treating clinician was blinded to which sessions were being reviewed for fidelity as well as the starting point for fidelity review within that session. Because the CPP software prompted the treating clinician to provide feedback on a predetermined schedule, fidelity checks were performed with reference to the CPP record for each session. For the selected sessions and trials, the reviewing clinician completed a checklist to verify the following aspects of the study design: (1) VAB was provided or not provided as indicated; (2) a model of the target was provided as indicated; (3) qualitative knowledge of performance (KP) feedback, such as articulation cues, visual-acoustic display cues, and verbal models, was provided as indicated; (4) knowledge of response feedback (i.e., correct or incorrect) was provided as indicated; (5) the clinician elicited a self-rating from the participant as prompted. Fidelity was over 99% for all parameters, reflecting both the experienced nature of the study clinicians and the high level of support provided by the CPP software. Further details are provided in Table 3.

Table 3.

Average fidelity of treating clinician’s adherence to CPP software prompts for biofeedback presentation, model prompts, knowledge of performance (KP) with articulation cues, KP with visual acoustic biofeedback (VAB), KP with a verbal model, knowledge of results (KR) feedback, and self-evaluation prompts, as indicated by study design.

	Biofeedback Presentation	Model prompt	KP Artic Cue	KP VAB Display	KP Model	KR	Self-Evaluation
Average	100.00%	100.00%	99.57%	99.29%	99.93%	99.93%	100.00%
SD	0	0	0.84	1.65	0.38	0.38	0

Perceptual Measurement of Treatment Progress

To quantify progress over the course of treatment, the present study obtained perceptual accuracy ratings for /ɹ/ sounds in untreated words administered during the baseline and maintenance phases. Probes were scored in real time by the treating speech-language pathologist, but because of the potential for these unblinded ratings to reflect listener bias, all probes were re-scored by listeners external to the study. Recordings of word probe measures administered before and after treatment were rated in a blinded fashion by untrained listeners recruited using the Amazon Mechanical Turk (AMT) crowdsourcing platform. In crowdsourcing, members of the general public perform a variety of online tasks for monetary compensation. Crowdsourcing has been used in a variety of behavioral studies, including to obtain ratings of clinical speech samples (Nightingale et al., 2020; Sescleifer et al., 2018). Although ratings from untrained individuals employing their personal equipment are more variable than those of trained listeners in a laboratory environment, validation research has indicated that this variability can be mitigated by collecting a greater number of ratings than is typical for lab-based investigations. For example, McAllister Byun et al. (2015) reported that ratings aggregated across nine untrained listeners recruited via AMT were comparable to ratings obtained across three expert listeners.

In the present study, ratings were acquired using the methodology suggested by McAllister Byun et al. (2015). Individual words were segmented out of each probe recording using Praat software (Boersma & Weenink, 2019). These word-level files were intensity-normalized and randomized into blocks of 200 files for presentation to listeners on AMT.¹ Each block was judged by nine unique listeners who rated the /ɹ/ sound in each word as “correct” or “incorrect.” Listeners were provided with an orthographic representation of the target word, but were not informed of the speaker’s identity or the time of elicitation (e.g., pre versus post). AMT raters were required to have IP addresses in the United States, report English as their primary language, and report no record of speech or hearing difficulties. Additionally, they were required to pass an initial eligibility task in which their ratings of 100 words containing /ɹ/ were compared to ratings given by expert listeners (McAllister Byun et al., 2015). At the beginning of each block, the participants were also required to pass a listening test to verify that they were using headphones (Woods et al., 2017). Furthermore, each group of files to be evaluated included 20 catch trials that had been deemed unequivocally correct or incorrect based on previous expert ratings. For their data to be included, blinded listeners needed to concur with the expert rater judgment on at least 80% of these tokens.

A total of 26 raters (19 women and 7 men) participated in the task. None of the raters reported a positive history of speech or hearing difficulty, and all reported speaking American English as their first language. Included raters had a mean age of 44.5 years (SD = 8.8, range = 33–60 years).

Interrater Agreement

Agreement across raters was quantified using the average proportion of raters who agreed with the modal rating for a given production (Spencer et al., in press). The mean rate of agreement across raters was 82.1%, in line with recommendations for a reasonable level of inter-rater agreement from previously published research (e.g., McAllister Byun et al., 2015).² A one-proportion z-test indicated that agreement across raters was significantly different from chance (50%), X-squared = 31058, p < 0.0001. Overall, the agreement across raters showed an adequate level of consistency.

Analyses

As noted above, the current analyses investigated the overall magnitude of participants’ responses to the combined treatment package without differentiating between traditional and biofeedback intervention types. Because baseline accuracy and treatment response for /ɹ/ may vary across phonetic contexts, analyses were conducted with all contexts pooled and with the separation of consonantal versus vocalic variants. Plots representing untrained listeners’ ratings of the accuracy of /ɹ/ words produced in the baseline and maintenance phases were visually inspected for each participant. To quantify the magnitude of change from pre- to post-treatment, effect sizes were computed using Busk and Serlin’s d₂ statistic (Beeson & Robey, 2006). The d₂ statistic pools standard deviations across the baseline and maintenance phases to minimize the number of cases in which an effect size cannot be calculated because of zero variance (which is most common during the pre-treatment baseline phase). Following previous research, we adopted 1.0, as the minimum threshold representing a clinically significant response to treatment (i.e., the difference between the post-treatment and pre-treatment means must be as least as large as the pooled standard deviation). Data visualization and analyses were performed in the R software environment (R Core Team, 2019) using the tidyverse family of packages (Wickham et al., 2019).

Results

Word probes: All data

The plots in Figure 3 depict the accuracy of the 50-word probe measures administered at baseline and maintenance intervals. Recall that these results do not differentiate between the effects of biofeedback and traditional treatment but only show the effect of the combined treatment package. The boxplots in Figure 3 represent the percentage of productions rated perceptually accurate when pooling across consonantal and vocalic variants of /ɹ/, as well as across the three sessions in each phase. Visual inspection of the boxplots suggested that five participants (3302, 3303, 6301, 6302, and 6303) demonstrated marked gains from pre- to post-treatment. Of the remaining two participants, one (3301) showed no change over the course of treatment, and one (3305) showed only a small increase in median perceptual accuracy.

Figure 3:

The graphs depict plotted accuracy on probe measures across baseline and maintenance intervals for visual inspection. Results are pooled over consonantal and vocalic contexts.

Effect size

The impressions from visual inspection of Figure 3 are corroborated by the mean level differences and standardized effect sizes reported in Table 4. The first columns show the mean and standard deviation of percent correct in the baseline period for each participant averaged across the three sessions. The following columns show the equivalent mean and standard deviation across the three maintenance sessions. The second to last column shows the difference between these two values, that is, the unstandardized effect size. This value was divided by the standard deviation pooled across baseline and maintenance sessions to arrive at the standardized effect size reported in the final column (Busk & Serlin’s d₂).

Table 4.

Mean level differences and standard deviations (SD) at baseline/maintenance timepoints and standardized effect size per participant.

Participant ID	Baseline mean	Baseline SD	Maintenance mean	Maintenance SD	Raw effect size	Effect size
3301	10.84	3.06	8.87	2.39	−1.97	−0.70
3302	3.13	2.18	24.42	1.49	21.30	10.99
3303	15.41	0.47	54.12	4.82	38.71	12.61
3305	54.42	6.98	64.60	0.81	10.18	1.87
6301	17.12	4.46	63.18	2.57	46.06	12.07
6302	33.02	5.30	80.16	4.34	47.14	9.54
6303	2.20	2.25	44.56	1.39	42.36	21.68

The raw change in percent correct from baseline to maintenance ranged from a minimum of −2 to a maximum of 47.1 percentage points. All participants, except 3301, achieved an overall standardized effect size exceeding 1.0, commonly treated as the minimum cutoff representing a clinically significant response to treatment (Maas & Farinella, 2012). Participant 3305, whose response to treatment was characterized as limited based on visual inspection, had a d₂ of 1.87, suggesting a small but clinically meaningful response. We will return to this participant below in our discussion of the differences across phonetic contexts. Several participants achieved large effect sizes up to a maximum d₂ of 21.7. (Note that the participant with the largest standardized effect size, 6303, did not demonstrate the largest raw effect size. This is because the standardized effect size is affected not only by the difference in means but also by the level of variability in the baseline and maintenance phases, with lower variability resulting in a higher standardized effect size.) Averaging standardized effect sizes across all participants yielded a mean d₂ of 9.7, suggesting that, on average, participants’ responses to the combined biofeedback and motor-based treatment package were positive and exceeded the minimum threshold for clinical significance.

Word probes: Separated by phonetic context

As a supplementary analysis, we examined plots and effect sizes separated by phonetic context. Specifically, we separated target words based on whether the target was a vocalic variant of /ɹ/ (defined here as syllabic or postvocalic /ɹ/) or a consonantal variant (i.e., /ɹ/ in onset position). In Figure 4, accuracy is plotted for consonantal variants at the top and vocalic variants at the bottom. Visual inspection of these graphs suggests that six of the seven participants made marked gains in perceptually rated accuracy for consonantal variants, while five of the seven participants made gains for vocalic variants. Effect sizes representing pre- to post-treatment changes in percent correct, separated by consonantal and vocalic context, are reported for all participants in Table 5.

Figure 4.

The graphs depict plotted accuracy on probe measures across baseline and maintenance intervals separated by consonantal and vocalic contexts.

Table 5.

Mean level differences, standard deviations (SD) and standardized effect size separated by consonantal and vocalic contexts.

Context	Subject	Baseline		Maintenance		Raw effect size	Effect size
		Mean	SD	Mean	SD
Consonantal	3301	26.07	7.59	22.42	6.57	−3.66	−0.51
	3302	6.64	5.10	48.01	1.07	41.38	10.33
	3303	26.55	3.04	44.99	4.92	18.44	4.73
	3305	16.81	3.19	47.45	4.30	30.64	8.34
	6301	32.21	10.12	68.16	8.66	35.95	3.76
	6302	51.43	5.85	71.69	3.24	20.25	4.07
	6303	4.00	4.72	57.78	5.55	53.78	10.62
Vocalic	3301	0.83	0.35	0.12	0.21	−0.71	−2.32
	3302	0.66	0.19	9.05	1.18	8.39	11.03
	3303	8.10	2.34	60.14	7.89	52.03	9.80
	3305	79.36	10.36	76.14	4.73	−3.22	−0.38
	6301	7.05	0.59	59.76	1.60	52.71	47.46
	6302	20.98	6.46	85.80	5.30	64.82	10.76
	6303	1.00	0.82	35.74	3.94	34.74	13.51

For consonantal variants, the raw change in percent correct from baseline to maintenance ranged from a minimum of −3.66 to a maximum of 53.78 percentage points. Except for Participant 3301, all participants showed standardized effect sizes well above the minimum clinically relevant standardized effect size of 1.0. However, the overall mean change in perceptually rated accuracy was smaller for consonantal variants (M = 5.92) than for the value pooled across phonetic contexts (M = 9.7). This may reflect the fact that for most participants, perceptually rated accuracy at baseline was higher for consonantal than vocalic variants of /ɹ/.

For vocalic variants, the raw change in percent correct from baseline to maintenance ranged from −0.71 to 64.82 percentage points, with five of the seven participants showing standardized effect sizes greater than the minimum threshold for clinical relevance.³ The mean score across the participants was 12.84. Overall, this supplementary analysis separated by consonantal versus vocalic variants reveals that the one participant whose overall response was marginal (3305) actually exhibited a clinically significant degree of improvement in consonantal /ɹ/ paired with a small negative change in accuracy for vocalic /ɹ/. The remaining six participants either responded to treatment for both variants (3302, 3303, 6301, 6302, and 6303) or demonstrated no meaningful response to treatment (3301).

Discussion

The results of the current investigation provide support for the efficacy of telepractice delivery of a treatment package combining VAB and motor-based treatment in children with RSSD. On average, participants made strong progress over the course of 20 sessions of telepractice treatment, showing an average raw improvement of 29.11 percentage points and a standardized effect size (d₂) of 9.7. This outcome compares favorably with a previous, similarly designed study of VAB and motor-based treatment in an in-person setting (McAllister Byun, 2017), which reported a mean effect size of 1.8. The notable difference in mean effect size may be related to the increased cumulative intervention index (CII; Warren, et al., 2007) in the current study compared with earlier work. While both study protocols reported equal dose frequency (number of treatment sessions per week) and total intervention duration (total time period of intervention), the present study elicited more trials per session (n = 150) than the previous study (n = 60). As a consequence, the CII was markedly higher in the present study, calculated at 3,000 trials compared to 1200 for McAllister Byun (2017). This highlights the importance of a sufficient dose in the treatment of RSSD (Hitchcock et al., 2019).

Similar reflections were reported by Peterson et al. (2022), which represents the closest counterpart of the current study in the previously published literature. That study also provided VAB and motor-based treatments for RSSD affecting /ɹ/ via telepractice in a design inspired by McAllister Byun (2017). VAB was delivered using the iOS application staRt (Speech Therapist’s App for “R” Treatment), rather than Sona-Match, as in the present study. The participants in Peterson et al. (2022) received a CII of 3,200 trials. Effect sizes based on untrained listeners’ perceptual ratings ranged from 5.25–67.55, with a mean of 27.64. The authors of that study pointed to the difference in CII as the likely driving force behind the difference in effect size relative to previous research (McAllister Byun, 2017). Overall, the present study agrees with Peterson et al. (2022) in reporting a positive, clinically significant average response to VAB and motor-based treatment for /ɹ/ delivered via telepractice. The results of the two studies differed in some respects (e.g., average standardized effect size), but these differences are difficult to interpret in light of the small number of participants enrolled in each study. Further research is needed to explore whether differences in software and hardware have a meaningful impact on the efficacy of VAB treatment delivered via telepractice.

Our study did include one participant who demonstrated no significant response to treatment (3301) and one participant whose response was marginal and conditioned by phonetic context. However, heterogeneity in treatment response, including non-responders, is typical in the literature on RSSD treatment. Previous studies using VAB have reported similar ranges of individual treatment responses, from no measurable difference to full mastery at the word level and beyond (McAllister Byun et al., 2016; McAllister Byun & Hitchcock, 2012; McAllister Byun & Campbell, 2016).

The present study included a supplementary analysis breaking down changes in perceptually rated accuracy across vocalic and consonantal variants of /ɹ/. Previous studies have reported conflicting results on the order of acquisition of different /ɹ/ allophones in both clinical contexts and typical development. Perception-based studies generally report earlier acquisition of consonantals than vocalic /ɹ/ (Curtis & Hardy 1959, Hoffman, Schuckers, & Daniloff 1980). On the other hand, studies based on acoustics tend to report earlier acquisition of vocalic /ɹ/ (McGowan et al., Klein et al., Magloughlin 2016). Researchers have argued that the apparent earlier acquisition of consonantal /ɹ/ may be attributable to differences in perceptual standards used to characterize /ɹ/ as correct or incorrect in different phonetic contexts. Specifically, Klein et al. (2012, 2013) argued that listeners may be more lenient in their perceptual ratings of consonantal /ɹ/ due to the shorter average duration of this variant. The present study is potentially consistent with previous findings that blinded listeners tend to give higher ratings to consonantal than vocalic /ɹ/ in the pre-treatment baseline phase. However, further study, including acoustic measurement, will be necessary to resolve this question.

Limitations

Several factors limit the strength of the conclusions that can be drawn from the findings reported herein. As indicated above, this paper reports only a preliminary analysis comparing accuracy before versus after administration of the complete treatment package. The larger study also collected probes before and after each session of treatment, which allowed us to measure increments of change within each session and compare these increments across VAB and motor-based treatment types. Because differences between conditions are likely to be small (McAllister Byun, 2017; Peterson et al., 2022), we considered it important to conduct these comparisons on the basis of acoustic measurements rather than perceptual ratings. Due to the time-intensive nature of acoustic measurement, this comparison will be reported in a separate paper. Thus, the main limitation of the present paper is that it does not allow us to weigh in on the relative efficacy of VAB versus motor-based treatment in the online setting. All but one of our participants had been receiving traditional articulatory treatment for /ɹ/ distortion for at least one year prior to enrollment in the present study. It is possible that participants’ strong average response to treatment in this study is a consequence of their exposure to VAB. It is also possible that exposure to motor-based treatment with evidence-based methods and high intensity as measured by CII, was the primary driving force behind their gains. Our ability to weigh in on the question of relative efficacy of treatment methods awaits the completion of the controlled comparison of VAB and motor-based treatment sessions. In the future, there is also a need for treatment comparisons using between-group designs to limit the likelihood of carryover between the conditions under study.

We also acknowledge that this study has other limitations. The small total number of participants makes it impossible to make strong inferences about generalizability to the broader population. Additionally, because our technological methods were still being refined as the study was launched during the COVID-19 pandemic, participants may have experienced intermittent challenges, such as lags in the response of visual signal to the participant’s voice, poor frequency resolution, and interruptions in internet connectivity. In addition, some participants exhibited behavioral challenges that could have been exacerbated by the fact that data collection occurred during COVID-19 lockdowns. It is necessary to follow up this pilot study with larger-scale studies of telepractice treatment for RSSD using a refined technological setup and investigating participants’ response in more typical (i.e., post-pandemic) conditions.

Conclusion

The present results support the hypothesis that a treatment package combining visual-acoustic biofeedback and motor-based treatment for rhotic distortions in children with RSSD can be effectively delivered via telepractice. There was evidence of a clinically meaningful response to the overall treatment package for 6 of the 7 enrolled participants. This study supports previous research in underscoring the importance of sufficient dosage in producing a response to treatment for RSSD. An analysis of the relative impact of traditional and biofeedback treatments will be undertaken in a separate paper. Overall, this body of research aims to inform clinical practice implementing visual-acoustic biofeedback and motor-based articulatory treatment for RSSD affecting /ɹ/ via telepractice.

Appendix A: Clinician instructions for audio and equipment setup for VAB telepractice treatment

Audio and Equipment Setup for Visual-Acoustic Biofeedback Telepractice Treatment
Computer	The participant and clinician must use a Windows or Mac operating system on a laptop or desktop computer. Cellphones are not permissible as they do not provide a large enough screen to effectively see the visual-acoustic biofeedback and clinician. Chromebooks are not permissible as the Zoom application cannot be downloaded on Chromebooks, which is necessary to have better access over Zoom audio controls.
Internet	To maximize the quality of the internet connection, the participant and clinician are required to plug the computer into the router using an ethernet cable instead of using wireless internet. If needed an ethernet cable will be provided.
Headset Microphone	The participant and clinician must use a Plantronics Blackwire C225 headset with microphone. The Blackwire C225 headset is a wired, binaural over-the-ear headset with a microphone frequency response of 100 Hz–10 kHz and hi-fi stereo output from 20 Hz–20 kHz. This headset has an appropriate frequency range to perceive the necessary frequency range required for /ɹ/.
Virtual Audio Mixer
Zoom Audio Settings	The Zoom application must be downloaded on both the clinician’s and participants’ computers in order to change the audio settings on Zoom to original sound. Zoom Audio Settings are found in the menu next to the microphone/mute function on Zoom. For the clinician’s Zoom Audio Settings, set the speaker to Voice Meeter and the microphone to the Plantronics Blackwire C225 headset (Jack Mic). For the participant’s Zoom Audio Settings, set the speaker and the microphone to the Plantronics Blackwire C225 headset (Jack Mic (Realtek (R) Audio)). To maximize audio and prevent sound drop-offs, in both the clinician’s and participant’s Zoom audio settings perform the following to configure the clinician’s and participant’s Zoom audio settings to use original sound: In the Audio Settings menu, select “Advanced” from the lower right-hand corner. In the “Advanced” Audio Settings Menu, check “Show in-meeting option to ‘Enable Original Sound’ from microphone.” Close the Audio Settings menu. Now a button that reads “Turn on original sound” will be at the top left-hand corner of the Zoom screen and will appear there for subsequent sessions. Prior to each session, click on the button so that the text changes to read “Turn off original sound”, which indicates that original sound is enabled. Both the clinician and participant are required to have original sound enabled during treatment sessions. Since the time of this study Zoom has updated the audio settings interface, enabling original sound is now labeled as “original sound for musicians” under Audio Profile in the Zoom Audio Settings.
Participant’s Probe Recording Device

Appendix B: Parent technology information handout for telepractice study

	Thank you for your participation in our virtual speech therapy study.
	What you will need prior to our first session:
	A desktop or laptop computer with Windows or Mac operating system. Tablets, cell phones and Chromebooks cannot be used for this study.
	Zoom application downloaded on your computer. If you don’t have Zoom software installed, go to https://zoom.us/ and follow the prompts to install it.
	Plug into the router. To maximize the quality of the internet connection, please plug the computer into the router using an ethernet cable instead of using WiFi. If you do not have an ethernet cable or are not able to connect to the router, please contact the study team to discuss options.
	Use headset with an in-line mic. The headset we sent you must be used for all evaluation and treatment sessions. We will setup the headset with you during our first call.
	Voice Recorder (for Windows) or Voice Memos (for Mac). During sessions, we may ask your child to record his/her voice to their computer and upload the recordings to our research team. Voice Recorder and Voice Memos come standard on most Windows and Mac operating systems. Please check that your computer has one of these recording devices prior to our first session. If not, please contact the study team to discuss options.
	Quiet space. Make sure the space your child will be in during sessions is as quiet as possible. If possible - no siblings/pets/TV in the background. Please minimize ventilation noise or any other machine noise.
	Be available for the first few minutes. It helps to have a parent available for the first few minutes before we start the assessments to help troubleshoot sound or video issues.
	If you have any questions, feel free to email us prior to your visit.

Appendix C: Clinician script for introducing VAB

Introduction to VAB with Vowels
Main Point(s)	Script
Show how to use visual display to make a better /ɹ/ sound.	“We’re going to use a computer program that makes a picture of your speech sounds. We’re going to use this to try and help you say a better /ɹ/ sound. Before we use the computer, we need to learn a little bit about what the computer image is telling us, so we know what to look for on the screen.”
Different sounds have different shapes on the LPC spectrum.
Moving tongue around in the mouth changes the tongue shape which moves the blue peaks around on the image.
Make clinician template of /i/ and demonstrate matching the template.
Load preselected speaker template match for /ɑ/, /i/, /u/ and encourage child to explore/match template shape.	“Now I want you to try. But our voices are different, right? So, if you try to match my red line, it may not line up too well. But we also have some lines that we traced from kids whose voices sound more like yours. We’ll have you try to match one of those lines instead.” <Select the preselected speaker template match for the /ɑ/ vowel.> “This line shows the shape of another kid’s /ɑ/ sound. I want you to say your best /ɑ/ sound into the microphone and see if you can make the bumps/peaks line up.” <Allow the client to play around until a satisfactory match with the template is reached. If client can’t achieve a match, select a different template.> <Load the /i/, /u/ templates from the same preselected speaker templates and encourage child to move tongue to match each template.>

NOTE: Introduction to VAB script reprinted with permission from Hitchcock et al. (2023).

Data available on request from the authors

The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.