Relative fundamental frequency in the assessment of primary muscle tension dysphonia

Vocal hyperfunction, characterized by excessive laryngeal muscle tension, is a condition associated with numerous voice disorders, including primary muscle tension dysphonia (pMTD). Primary MTD is disturbance occurring in the absence of structural or neurologic pathology. Vocal hyperfunction is postulated to be the proximal cause of the dysphonia in pMTD. A measure of relative fundamental frequency (RFF) has been proposed as an objective and noninvasive marker of vocal hyperfunction. This retrospective study examined the clinical utility of RFF in identifying and tracking changes in vocal hyperfunction before and following a course of manual circumlaryngeal therapy for pMTD patients. Because RFF is a time-based acoustic measure reliant on periodicity for accurate calculation, additional investigation into the influence of dysphonia severity on its validity across the severity spectrum was completed. RFF calculations were derived from pre- and posttreatment audio recordings from 111 females with pMTD and 20 vocally normal controls. Three voiced-voiceless stimuli (VCV tokens) were analyzed. Listener ratings of dysphonia severity were employed to determine (1) the relation of RFF measures to overall severity, and (2) the effects of dysphonia severity on the utility of RFF calculations. Multiple regression analyses demonstrated that RFF onset slope consistently varied as a function of group membership and therapy time conditions. Pearson Product-Moment Correlations showed a significant relationship between RFF onset cycle 1 values and listener ratings of dysphonia severity. In addition, regression analysis confirmed the influence of therapy condition and specific RFF cycles on dysphonia severity rating. Cumulatively, the analyses confirmed that RFF onset may be sensitive to predicting the presence and degree of vocal hyperfunction before and after therapy, and as an index of dysphonia severity. However, RFF could not be fully analyzed in many subjects, yielding a large quantity of missing data. Adjusted odds ratio estimates revealed that these unanalyzable data were related to phonetic context (token), group membership (pMTD vs. control) and severity level of dysphonia. Although RFF showed potential as an objective measure of vocal hyperfunction before and following voice therapy, the large number of unanalyzable samples (related to increased dysphonia severity especially in the pMTD group) represents a critical limitation.

RELATIVE FUNDAMENTAL FREQUENCY IN THE ASSESSMENT OF PRIMARY MUSCLE TENSION DYSPHONIA by Rebecca Ann Fetrow A thesis submitted to the faculty of The University of Utah in partial fulfillment of the requirements for the degree of Master of Science in Speech-Language Pathology Department of Communication Sciences and Disorders The University of Utah August 2015 Copyright © Rebecca Ann Fetrow 2015 All Rights Reserved T h e U n i v e r s i t y o f U t a h G r a d u a t e S c h o o l STATEMENT OF THESIS APPROVAL The thesis of Rebecca Ann Fetrow has been approved by the following supervisory committee members: , Chair 4/28/15 Date Approved Bruce Smit h , Member 4/28/15 Date Approved Nelson Roy , Member 4/28/15 Date Approved and by Michael Blomgren , Chair/Dean of the Department/College/School of Communication Sciences and Disorders and by David B. Kieda, Dean of The Graduate School. Michael Blomgren ABSTRACT Vocal hyperfunction, characterized by excessive laryngeal muscle tension, is a condition associated with numerous voice disorders, including primary muscle tension dysphonia (pMTD). Primary MTD is disturbance occurring in the absence of structural or neurologic pathology. Vocal hyperfunction is postulated to be the proximal cause of the dysphonia in pMTD. A measure of relative fundamental frequency (RFF) has been proposed as an objective and noninvasive marker of vocal hyperfunction. This retrospective study examined the clinical utility of RFF in identifying and tracking changes in vocal hyperfunction before and following a course of manual circumlaryngeal therapy for pMTD patients. Because RFF is a time-based acoustic measure reliant on periodicity for accurate calculation, additional investigation into the influence of dysphonia severity on its validity across the severity spectrum was completed. RFF calculations were derived from pre- and posttreatment audio recordings from 111 females with pMTD and 20 vocally normal controls. Three voiced-voiceless stimuli (VCV tokens) were analyzed. Listener ratings of dysphonia severity were employed to determine (1) the relation of RFF measures to overall severity, and (2) the effects of dysphonia severity on the utility of RFF calculations. Multiple regression analyses demonstrated that RFF onset slope consistently varied as a function of group membership and therapy time conditions. Pearson Product- Moment Correlations showed a significant relationship between RFF onset cycle 1 values and listener ratings of dysphonia severity. In addition, regression analysis confirmed the influence of therapy condition and specific RFF cycles on dysphonia severity rating. Cumulatively, the analyses confirmed that RFF onset may be sensitive to predicting the presence and degree of vocal hyperfunction before and after therapy, and as an index of dysphonia severity. However, RFF could not be fully analyzed in many subjects, yielding a large quantity of missing data. Adjusted odds ratio estimates revealed that these unanalyzable data were related to phonetic context (token), group membership (pMTD vs. control) and severity level of dysphonia. Although RFF showed potential as an objective measure of vocal hyperfunction before and following voice therapy, the large number of unanalyzable samples (related to increased dysphonia severity especially in the pMTD group) represents a critical limitation. iv vi TABLE OF CONTENTS ABSTRACT……………………………...…………………………………………...….iii LIST OF TABLES………...…………………………………………………………..…vii LIST OF FIGURES……………………………………………………………………..viii ACKNOWLEDG􀀨􂡍MENTS………………………………………………….…….……􀀑􁅩ix INTRODUCTION………………………………………………………………………...1 Vocal Hyperfunction and Primary Muscle Tension Dysphonia (pMTD)……..….1 Relative Fundamental Frequency (RFF): A Marker of Vocal Hyperfunction?......2 Research Questions…………………………………………………………….....9 METHODS………………………………………………………………………………11 Participants…………………………………………………………………….....11 Speech Task and Recording……………………………………………………...12 Listener Perceptual Ratings……………………………………………………...13 RFF Calculation………………………………………………………………….14 Re-measurement Reliability……………………………………………………...16 RESULTS……………………………………………………………………………...…18 VCV Token Analysis……………………………………………………………..18 RFF Differences across Group and Time Conditions…………………………....19 RFF and Dysphonia Severity………………………………………………….....26 Missing/Unanalyzable RFF Data………………………………………………...29 DISCUSSION…………………………………...……………………………………….34 RFF Onset Values across Group and Time Conditions…………………………..34 Correlations between RFF and Listener Perceptual Ratings.................................37 Dysphonia Severity and RFF Calculations……………………………………....38 Causes of Missing/Unanalyzable Data…………………………………………..40 Automation of RFF Calculation………………………………………………….44 CONCLUSIONS…………………………………...……………………………………46 APPENDIX…………………………………………………………………………..…..48 REFERENCES……………………...………………………………………...…………51 vi vi LIST OF TABLES Tables 1. Participant Ages by Group………………………………………………………..…...12 2. ANOVA for Cycle and Token Effects……………………………………………........19 3. Multiple Regression Analysis for Cycle and Group Effects………………………..…21 4. Multiple Regression Analysis for Group, Time, and Age Effects…………………......26 5. RFF Means, Standard Deviations, Ranges, and Numbers by Token……..………..….27 6. Correlations of RFF 􀀹􃥡alue and 􀀶􃙥everity…………………..…………………………􀀑􁄲28 7. Regression Analysis for RFF 􀀦􂙹ycle and Time Effects on Severity……..…………….29 8. Number of Excluded RFF Patterns (out of 262 possible)………..…………………....30 9. Mean Number of Analyzable Tokens………………..………………………………...30 10. Odds Ratio Estimates……………………………………………………………..….31 11. Missing Data by Severity Quartile…………………………………………..……….32 LIST OF FIGURES Figures 1. Distribution of Severity by Group and Time…………………..…………………...…15 2. Waveform in the Time Domain…….…………………….……………………………16 3. Token 1 RFF Mean Values by Group and Time…………………..……………...........23 4. Token 2 RFF Mean Values by Group and Time……………………..………………...24 5. Token 3 RFF Mean Values by Group and Time……………………..…………...........25 ACKNOWLEDGEMENTS I would like to express my gratitude to Dr. Nelson Roy for his mentorship, sharing of knowledge, and encouragement over the course of this project. His excellent advice and support resonated in multiple facets of my educational experience, both creating a positive learning environment and laying the foundation for my future career in voice disorders. Especial thanks also to Dr. Ray Merrill, whose statistical analyses approached genius as he relentlessly tackled this challenging set of data from many angles. My appreciation to Dr. Cara Stepp, Dr. Stephanie Lien, and their entire research team for their previous studies in the area of relative fundamental frequency and for encouragement of my own project, Dr. Christopher Dromey and his graduate students for providing the listener-perceptual ratings used in this study, and Hannah Allen for contributing interrater reliability measures. My sincerest thanks and recognition to Dr. Michael Blomgren and Dr. Bruce Smith for being part of my graduate thesis committee. Finally, much love to my friends and family for their ceaseless support over the past two years. It has been quite a journey and I am grateful to all who helped me along the way. INTRODUCTION Vocal Hyperfunction and Primary Muscle Tension Dysphonia Without exception, contemporary voice texts cite excessive or poorly regulated activity of the intrinsic and extrinsic laryngeal musculature as contributing to a host of vocal pathologies (Aronson, 1990; Boone & McFarlane, 2000; Colton & Casper, 2006; Morrison & Rammage, 1994; Stemple, 2000), including a class of disorders referred to as hyperfunctional or musculoskeletal tension voice disorders (Hillman, Holmberg, Perkell, Walsh, & Vaughn, 1989). Several definitions of vocal hyperfunction exist, but a recurrent feature in almost all descriptions includes excessive laryngeal musculoskeletal activity, force, or tension accompanying voice production (Roy, 2008). While vocal hyperfunction has been implicated in a variety of voice disorders (such as nodules and polyps), it has been most prominently associated with a disorder known as primary muscle tension dysphonia (pMTD). Primary MTD is a voice disturbance that occurs in the absence of visible structural or neurological vocal fold pathology, with vocal hyperfunction (i.e., in the form of excessive or dysregulated activity of intrinsic and extrinsic laryngeal muscles) frequently cited as the proximal cause of pMTD (Aronson, 1990; Boone & McFarlane, 2000; Colton & Casper, 2006; Morrison & Rammage, 1994; Stemple, 2000). Primary MTD accounts for 10% - 40% of cases seen in multidisciplinary voice disorder clinics (Roy, 2003). Despite the prevalence of vocal hyperfunction syndromes like pMTD, diagnosis 2 and evaluation of vocal hyperfunction remains reliant upon patient histories, physical examinations, and qualitative assessments of auditory-perceptual cues by experienced clinicians, which are inherently subjective (Stepp, Hillman, & Heaton, 2010). The effectiveness of behavioral interventions for vocal hyperfunction can, therefore, only be assessed by subjective measures like listener perception, palpation, patient report, visual examination of the neck and connecting tissues, and laryngovideostroboscopy (Stepp, Merchant, Heaton, & Hillman, 2011). A therapist's ability to accurately, reliably, and objectively quantify change resulting from voice therapy techniques remains compromised. Of the subjective measures currently used clinically, auditory perception of dysphonia severity, vocal effort (VE), or strain is considered the "gold standard" for estimating vocal hyperfunction across multiple voice disorders (Stepp et al., 2011). Resulting ratings are dependent upon the individual evaluator's experience and training, and may not be valid or comparable. Despite its importance in the treatment strategies for many common voice problems, no objective clinical measure has been established to accurately estimate the presence and severity of vocal hyperfunction (Stepp et al., 2010). The need for quantifiable measures has led researchers to explore the areas of acoustics, aerodynamics, and electromyographic signals in search of a more reliable and valid means of determining vocal hyperfunction (Stepp et al., 2011). Relative Fundamental Frequency: A Marker of Vocal Hyperfunction? In an attempt to identify an acoustic measure that would highly correlate with vocal effort ratings, several recent studies examined the changes in relative fundamental frequency (RFF) surrounding vowel-voiceless consonant-vowel (VCV) productions. 3 These studies have offered support for RFF as a potentially inexpensive and noninvasive marker of vocal hyperfunction (Stepp et al., 2010; Stepp et al., 2011). RFF is defined as the "fundamental frequency of the cycles immediately before and after production of a voiceless consonant, normalized by the 'steady-state' fundamental frequencies of the voicing preceding and subsequent to the consonant" (Stepp et al., 2010). Normalization describes a process whereby all fundamental frequencies measured are "converted to semitones (ST) relative to the points in the voicing furthest from the voiceless consonant (Cycle 1 in voicing offset and Cycle 10 in onset)" (Stepp et al., 2010, pg. 1222-1223). Individual variations as well as gender and age differences with respect to fundamental frequency make normalization a critical aspect of this measure. Men generally speak at a lower fundamental frequency than do women, and there can also be great variability among females and age groups (Robb & Smith, 2002). By normalizing semitones (ST) in relation to "steady-state" fundamental frequencies, comparison across multiple speakers is made possible (Stepp et al., 2010). Fundamental frequency (F0) is dependent upon the rate at which the vocal folds vibrate in response to their length, mass, applied degree of tension (Van Den Berg, 1958) and paralaryngeal or subglottal air pressure (Titze, 1989). Both conscious and unconscious rate variations are possible with changes to underlying laryngeal muscle tension across multiple time intervals, including the sentence level or intonation contour, the word and syllable or lexical stress level, and across individual phonemic segments (Crystal, 1969). Because of augmented vocal fold tension and increased air pressure used to inhibit vibration during voiceless consonant production, vowels following voiceless consonants have higher onset fundamental frequencies than those following voiced 4 consonants (House & Fairbanks, 1953; Stevens, 1977). The fluctuations of F0 immediately surrounding these individual phonemes are laryngeal adjustments required to execute necessary articulatory and co-articulatory events, but do not contribute to the meaning of an utterance (Löfqvist et al., 1995). As they are not consciously controlled, these short time intervals provide an ideal context in which to evaluate the laryngeal behaviors of healthy and unhealthy phonation (Watson, 1998). Relative fundamental frequency capitalizes upon these principles by evaluating brief time intervals at the phonemic level of vowels on either side of a voiceless consonant: the VCV pattern. Specifically, RFF during both vowel offset and onset has been shown to follow characteristic patterns that differ between healthy and disordered voices (Robb & Smith, 2002). During healthy phonation for speakers of all ages, increases in RFF are observable during vowel onset following an unvoiced consonant (Goberman & Blomgren, 2008; Robb & Smith, 2002). RFF during vowel offset into a voiceless consonant is typically stable in young individuals, while older speakers exhibit a marginal decrease in RFF (Goberman & Blomgren, 2008; Watson, 1998). Vocal hyperfunction results presumably from excessive amounts of laryngeal tension. This tension potentially interferes to a measurable degree with normal RFF values, as baseline levels of laryngeal tension may be increased to the point where typical short-term variations in patterns of voicing and devoicing are no longer possible (Stepp et al., 2010). A review of the literature demonstrates an expanding body of evidence in support of the clinical potential of RFF as an acoustic measure of vocal hyperfunction, although limitations persist and research is still in its early stages. Stepp, Hillman, and Heaton (2010) originally hypothesized that RFF values 5 surrounding voiceless consonants would be depressed in patients with vocal hyperfunction. Their retrospective study included women presenting with vocal nodules, vocal polyps, or pMTD. While the presence of vocal lesions was not found to impact RFF pre- and postsurgery, there was a significant lowering of the offset and onset RFF values across all groups presenting with vocal disorders as compared to normal controls. It was concluded that "altered offset and onset RFF in patients with hyperfunction-related voice disorders can be interpreted as a by-product of heightened levels of laryngeal muscle tension. Measurement of RFF during voice offset and onset has potential for use as a simple, noninvasive measure of vocal hyperfunction" (Stepp et al., 2010, p. 1220). Numerous subsequent studies have attempted to assess RFF's usefulness in a variety of ways. For instance, RFF may have the potential to quantify vocal function changes directly resulting from voice therapy. Stepp, Merchant, Hillman, and Heaton (2011) sought to determine the effect of a single course of voice treatment on RFF in vowel-voiceless consonant-vowel tokens. Pre- and posttherapy RFF calculations were evaluated for each of 16 female patients diagnosed with either bilateral vocal fold nodules or MTD. It was anticipated that the lowered RFF values characteristic of vocal hyperfunction would normalize following treatment. Indeed, RFF offset and onset values following therapy increased to more closely approximate typical measurements associated with healthy phonation. This normalizing effect may be the direct result of the voice therapy's successful diminishing of baseline levels of laryngeal muscle tension, and suggests potential for RFF to be used clinically as a treatment outcomes measure. As previously described, auditory-perceptual ratings remain the "gold standard" to assess vocal hyperfunction and its severity across multiple voice disorders. Stepp, Sawin, 6 and Eadie (2012), in an attempt to correlate an acoustic measure "with the perceptual attributes of vocal effort (VE) or strain" (pg. 1888), compared RFF to ratings of both overall severity (OS) and VE. Twelve listeners evaluated the speech samples of 30 hyperfunctional and 10 control voices using rating scales of OS and VE. A near perfect correlation was shown between VE and OS scores. When compared with manually calculated RFF, weak negative correlations existed between both perceptual measures and RFF offset cycle 10 values, while no correlations were found with RFF onset cycle 1 values. The disordered voices in this study were scored perceptually to have mild to moderate levels of dysphonia severity, and even to sometimes fall within normal limits as compared with healthy controls. While overlap in VE scores existed for a large portion of the control and disordered groups, mean RFF values demonstrated distinct differences in the patterns between the groups. This finding suggested "that RFF more accurately classifies the presence of a voice disorder than does severity of voice quality or VE" (p. 1893), making it a potentially superior diagnostic tool to clinical judgment and auditory-perceptual ratings alone. Several additional studies were completed to better explore the potential relationship between VE/OS and RFF. Stepp and colleagues (2012) determined that onset RFF, not offset RFF (as shown prior) was related to VE ratings in a sample of patients with adductor spasmodic dysphonia, making onset RFF potentially viable as an outcome measure for similar populations (Eadie & Stepp, 2013). In another study, the potential for RFF to be used to monitor changes in VE both across and within individuals was examined (Lien et al., 2015). Instead of examining a disordered voice group, 12 healthy individuals were asked to produce a single VCV token many times, purposefully 7 changing their VE in specific ways during the speech task. While correlations across speakers for both RFF and an aerodynamic measure of VE were weak (R2=.06-.26), they were much stronger within speakers on average (R2=.45-.56). Similar research with voice disordered groups could further demonstrate potential for RFF to be used as a treatment outcomes measure. Eadie and Stepp (2013) used the voiceless consonant laden sentence, "He saw half a shape mystically cross a path about fifty or sixty steps from his sister Kathy's house," for RFF measurements, and a total of 9 voiced-voiceless-voiced (VCV) tokens or RFF productions were extracted for analysis. It was found that when 6 or more tokens were viable for RFF measure, "the strength of the relationships between the RFF of onset cycle 1 and the perceptual measures" were optimally stable (pg. 173). While an important finding relevant to future research and clinical application, the investigators recognized that if 6 tokens were required for accurate RFF estimation, the time requirements inherent to manual calculation could be prohibitive for a relatively inexperienced clinician working in a demanding clinical setting. For this reason, the development of an automation process for RFF calculation was explored (Lien & Stepp, 2013). An algorithm to automate RFF estimation was tested on the recordings of 12 healthy controls and 12 participants with Parkinson's disease (PD). RFF has previously been examined in the PD population, finding that RFF is lower in patients affected by PD progression (Stepp, 2013). Interfacing with Praat acoustic analysis software, "pulse timings that occurred within a time window specified by the MATLAB algorithm" were determined, and adjustments for gender differences in pitch were made (Lien & Stepp, 2013, pg. 2137). Criteria for rejecting VCV tokens were similar to those for manual 8 calculation. Similarities between automated and manually calculated RFF for isolated VCV tokens (i.e., voiced-voiceless pairings that were elicited as isolated syllables as opposed to in spontaneous or connected speech, such as /isi/ or /apa/) were encouraging. However, inconsistencies with cycle detection and the current inability to apply the proposed algorithm to running speech are limitations that require further investigation (Lien & Stepp, 2013). Several additional factors have been raised as perhaps being especially important to accurately estimating RFF in healthy and disordered voice populations. These are phonetic context (i.e., the specific phonemes associated with VCV tokens, as well as isolated or running speech contexts) and the type of recording signal used for RFF extraction. A comparison of accelerometer, low-pass filtered, and unprocessed microphone signals revealed that small significant effect sizes were found in RFF estimation in a group of healthy voices, but the overall patterning of RFF was similar across signal type (Lien & Stepp, 2014). Similar research to explore the effects of signal type on RFF in disordered voices may help to establish the need for standardization of recording procedure in a clinical setting. More relevant to the current study was an investigation into the possible effects of phonetic context on RFF calculation towards the development of a set of standardized stimuli for future studies and clinical practice (Lien et al., 2014). The interactions among aerodynamics, vocal fold motion, and muscular tension during voicing may vary depending on phonetic context. This may significantly influence RFF estimation, meaning that some voiced-voiceless stimuli (i.e., VCV tokens) may be more optimal or consistent for RFF calculation than others. In healthy voices, it was found that stimuli 9 containing the voiceless consonants /f/ and /ʃ / had the lowest standard deviations and were, therefore, recommended most highly for RFF analysis. Similar studies in the voice disordered population, as well as research examining the voiced portions of the stimuli and isolated syllable versus running speech contexts, are important considerations for future research. Research Questions The current investigation was designed to address several research questions. First, this study sought to evaluate whether RFF was a sensitive index of vocal hyperfunction, both before and following a successful course of voice therapy. It could then be determined whether RFF reliably tracked changes following successful management of pMTD. Previous studies have examined vocal hyperfunction irrespective of etiology. By analyzing RFF in pure cases of pMTD only, a condition that occurs in the absence of structural or neurological pathology and is believed to result from excess laryngeal tension (i.e., vocal hyperfunction), one can presumably assess the pure effects of vocal hyperfunction uncomplicated by other possible confounds. Second, this study explored the influence of dysphonia severity on RFF's validity as a measure of hyperfunction. More severely dysphonic voices tend towards aperiodicity, which interferes with accurate tracking of fundamental frequency (Bielamowicz, 1996). Furthermore, increasing aperiodicity in an acoustic sound pressure waveform increases the difficulty in accurately identifying cycle boundaries, both visually and for automated acoustic software analysis programs (Titze, 1995). It was, therefore, speculated that a dysphonia severity threshold might exist, beyond which RFF would fail to be viable, and that accurate extraction (i.e., measurement) of RFF would be 10 related to overall severity of dysphonia. The existing literature that has evaluated the utility of RFF has included principally patients with mild-moderate dysphonia, and the impact of dysphonia severity on RFF measurement has been relatively understudied. It is well known that pMTD spans the severity spectrum, and includes some of the most severely disordered voices encountered clinically (Roy, 2003). Thus, pMTD represents an ideal clinical population to evaluate the robustness of RFF as a potential clinical tool to identify and quantify vocal hyperfunction across the entire severity spectrum. METHODS Participants An existing database/corpus of pre- and posttreatment audio-recordings of patients with pMTD was employed in this study. These recordings were originally collected by Dr. Nelson Roy during clinical practice. A diagnosis of pure pMTD, based on case history, auditory-perceptual evaluations of the voice, endoscopy, and an assessment of musculoskeletal tension (Roy & Bless, 1998), was made for 157 patients by a team composed of a laryngologist and a certified speech-language pathologist. The recordings were reviewed sequentially from first to last for inclusion criteria, resulting in a final database of 111 individuals. Inclusion in the database required positive change as a result of behavioral voice therapy, and all therapy sessions included manual circumlaryngeal reposturing, massage, or a combination of both techniques administered by Dr. Nelson Roy. Voice therapy sought to stimulate improvements in the quality of phonation during connected speech by interfering with habituated muscle misuse patterns, reducing laryngeal tension, improving laryngeal posture, and balancing muscular activity (Roy et al., 2009). Women who responded positively to a single extended session of voice therapy (rated by both patient and clinician) and who had a complete and analyzable data set were included in this database, regardless of severity. This database has been used in a number of previous studies (Dromey et al., 2008; Roy et al., 2009). For comparison purposes, an existing dataset of vocally normal control subjects 12 (N= 20 females) was used to examine potential practice effects (i.e., possible effects related to reading the sample twice). Control subjects were all students at Brigham Young University with no history of hearing, speech, voice, or language disorders and were provided courtesy of Dr. Christopher Dromey. Because the controls were not precisely matched by age to the pMTD subjects in this study, age served as a covariate in later statistical analysis. A breakdown of the participants' ages is provided in Table 1. In total, the final database included recordings from 111 female participants with pMTD (M=46.17 years, SD=13.69) and 20 female control speakers (M=23.35 years, SD=7.01). The University of Utah Institutional Review Board granted permission to use these voice samples and negated the need to request new permission from patients in all subsequent studies (IRB 00078937). Speech Task and Recording Pretreatment and posttreatment readings of the second and third sentences from The Rainbow Passage (Fairbanks, 1960) were audio recorded for all pMTD patients at a comfortable pitch and loudness level. Recordings of the same sentences read by the vocally normal controls were made approximately 1 hour apart as this represented the time span of a typical therapy session. Audio recordings were digitized offline at 25 kHz with a Computerized Speech Lab system (Kay Elemetrics; Lincoln Park, NJ). While the patient recordings were collected over a span of several years and recording equipment Table 1. Participant Ages by Group. Group Number of Participants Mean Age (years) Standard Deviation Range (years) pMTD 111 46.17 13.69 12-79 Control 20 23.35 7.01 19-52 13 across patients varied, each individual subject's voice was captured using identical equipment on the same day and in the same location. Therefore, each subject acted as her own control and differences within subjects could be viewed as treatment effects, not as recording inconsistencies (Roy et al., 2009). Voice samples were recorded as part of a routine clinical evaluation and were not originally intended to be used for acoustic analysis. While this limited the number and types of stimuli tokens suitable for this study to the first two sentences of the selected reading passage, it also added an element of "naturalness" to the task as compared with isolated VCV repetitions that do not emulate patterns typical of connected speech (Roy et al., 2009). Listener Perceptual Ratings In clinical circles, the extent of vocal hyperfunction is believed by many to be reflected in the severity of the voice quality disturbance, with increasing levels of hyperfunction contributing to increasing levels of dysphonia severity. In order to examine RFF in relation to possible vocal hyperfunction, listener ratings of dysphonia severity were used from a previous study (Dromey et al., 2008) that employed these identical audio recordings. In that study, five master's level students majoring in communication disorders were recruited as listeners to provide auditory-perceptual scores for the database of recorded pMTD and control voice samples. Listeners evaluated the same database of fully randomized audio samples described previously, presented to each listener in an identical order. Each student listened to all samples via headphones at a self-determined comfortable loudness level within a quiet laboratory. A custom MATLAB routine was developed to rate each sample as it was heard. Seated at a computer, each individual 14 used a computer mouse to move a cursor along a visual analog scale labeled "normal" at one end and "profoundly abnormal" at the other. Based on the position of the cursor, the computer software documented a score of 0-100 for each sample, with larger numbers representing increased dysphonia severity. Thirty-eight samples were repeated at the end of the listening session for the purpose of evaluating intrarater reliability. Pearson correlations between the first and second scores for the 38 repeated voice samples revealed a range of .95-.98 (M=.97) for individual raters, indicating excellent reliability within each listener. Interrater reliability was assessed using the intraclass correlation coefficient (ICC). An average-measure ICC of .97, with a single-measure ICC at .88, [F(261, 1044)=38.942, p<.001] confirmed acceptable interrater reliability (Dromey et al., 2008). A distribution of severity ratings for the pMTD and control groups is displayed in Figure 1. Severity was not found to be significantly associated with age (F=1.88, p=0.17), but it was significantly associated with both groups - control or pMTD (F=240.92, p˂.001) and time - pretherapy or posttherapy (F=785.10, p˂.001). RFF Calculation Five VCV tokens from the identical second sentence of The Rainbow Passage, which was employed in the listener rating task, were selected for RFF analysis: "beautiful" (/ə/-/f/-/əl/), "the shape" (/ə/-/ʃ/-/e/), "shape of" (/e/-/p/-/ə/), "apparently" (/ə/- /p/-/ɝ/), and "the horizon" (/ə/-/h/-/ɚ/). Procedures for RFF calculation closely followed those described in previous research, (see Stepp, Merchant, Heaton, and Hillman, 2011). Waveforms (in the time domain) for each of the voice sample recordings were displayed using Praat acoustic analysis software (Boersma & Weenink, 2014) at a zoom level of 0.3-0.5 seconds. In order to determine RFF, the investigator measured the 10 cycles of 15 Figure 1. Distribution of Severity by Group and Time. vibration of both vowel offset and onset surrounding a voiceless consonant using the software's pulse function. An illustration of a typical waveform is displayed in Figure 2 with RFF offset and onset cycles labeled, as well as phoneme locations. The instantaneous fundamental frequency was then calculated as the inverse of each period, and all frequencies were normalized by converting them to semitones "relative" to the first cycle in the 10 cycles prior to voicing offset and the final cycle in the 10 cycles following voicing onset using the following formula (Baken, 1987, pg. 127): ST=39.86 × log10(f/fref). Offset and onset RFF calculations were analyzed for all 5 VCV tokens in 262 N Mean SD Mean SD pMTD 111 67.96 27.81 14.05 15.45 Controls 20 1.19 2.62 1.31 2.84 16 Figure 2. Waveform in the Time Domain audio samples across the 131 subjects. Primary MTD subjects accounted for 222 of the audio recordings, as both pre- and posttherapy conditions were recorded for each of the 111 pMTD cases. The 20 control speakers, recorded twice, added an additional 40 recordings. Remeasurement Reliability In order to assess intrarater RFF measurement reliability, the same 38 samples randomly selected from the 262 audio samples (approximately 15% of the total sample) in the listening task were reanalyzed. The primary investigator, who was blinded to the identity of the sample, remeasured RFF. To assess interrater reliability, 10% of the recordings were randomly selected from the database and a second examiner measured RFF. The second examiner, who was trained by the primary investigator, was also blinded to the identity of the sample. Because the data were not normally distributed, intra- and interrater measurement reliability was assessed using Spearman correlation coefficients (i.e., rs). Intrarater reliability was determined to be good as measured with a 17 Spearman correlation coefficient (rs=0.734, p=˂.001). Interrater reliability was also good evidenced by a Spearman correlation coefficient (rs=0.744, p= ˂.001). RESULTS VCV Token Analysis While RFF was originally calculated for 5 VCV tokens in total, only 3 of those tokens proved viable for statistical analysis: "beautiful" (/ə /-/f/-/əl/), "the shape" (/ə/-/ʃ/- /e/), "shape of" (/e/-/p/-/ə/). An extremely high proportion of missing data for the remaining 2 tokens, "apparently" (/ə/-/p/-/ɝ/) and "the horizon" (/ə/-/h/-/ɚ/), was judged as unacceptable and precluded these tokens from analysis. For instance, token 4, "apparently," was one example of a vowel offset duration that was consistently too brief for accurate extraction of 10 vibratory cycles. Furthermore, the word "apparently" was not read by 36 subjects in the pMTD group, possibly as a result of an erroneous copy of The Rainbow Passage in which the word was omitted. Because usable offset data for this token were only available for 22.9% of the entire database, it was decided to eliminate the token from final analysis. Likewise, token 5, "the horizon," also represented a VCV context with a high proportion of missing data. The "h" phoneme in this phonetic context was only devoiced by speakers in approximately 25% of samples, meaning there was no distinguishable period of voiceless production for the phoneme and subsequently no discernible offset and onset of the vowels surrounding the "h" in 75% of samples. Token 5 was therefore also eliminated from the set, leaving tokens 1, 2, and 3 for final analysis: "beautiful" (/ə/-/f/-/əl/), "the shape" (/ə/-/ʃ/-/e/), "shape of" (/e/-/p/-/ə/). Previous studies have shown no significant differences among various voiceless 19 consonant productions, allowing for averaging across tokens (i.e., phonetic contexts) to produce a single mean RFF patterns of offset and onset to be used for analysis (Stepp et. al, 2010; 2011). However, prior to aggregating across tokens in this study, a two-way ANOVA was computed to examine the effects of cycle, token, and cycle by token interaction for each offset and onset RFF pattern in each unique group and therapy time condition (Table 2). Significant cycle by token interaction effects for both pMTD and control groups precluded pooling of the RFF data across tokens. In other words, the RFF values among tokens 1, 2, and 3 behaved differently enough that aggregating data into a single mean pattern of offset RFF and single mean pattern of onset RFF average for each subject was deemed inappropriate. Each of the 3 tokens used in this study was therefore statistically analyzed individually and results are presented and discussed on a token by token basis. RFF Differences across Group and Time Conditions A primary objective of this study was to determine if and how RFF patterns in pMTD varied from patterns observed in the control group, as well as the extent to which they varied between pretherapy and posttherapy conditions (time 1 vs time 2). If one Table 2. ANOVA for Cycle and Token Effects. Conditions Cycle Token Cycle*Token Time 1 RFF Offset pMTD F=49.35, p=˂.001 F=1.94, p=0.166 F=4.99, p=0.016 Control F=36.43, p=˂.001 F=0.15, p=0.861 F=1.82, p=0.184 RFF Onset pMTD F=141.69, p=˂.001 F=16.25, p=˂.001 F=8.14, p=0.020 Control F=261.18, p=˂.001 F=9.89, p=0.007 F=4.22, p=0.027 Time 2 RFF Offset pMTD F=104.64, p=˂.001 F=3.56, p=0.044 F=60.33, p=˂.001 Control F=204.51, p=˂.001 F=2.80, p=0.080 F=23.11, p=˂.001 RFF Onset pMTD F=393.88, p=˂.001 F=9.81, p=0.008 F=6.33, p=0.006 Control F=351.25, p=˂.001 F=11.66, p=0.003 F=7.33, p=0.003 Significance is bolded. Time 1 = Pretherapy for the pMTD group, Time 2 = Posttherapy for the pMTD group. 20 accepts that in patients with pMTD, the primary focus of voice therapy is to minimize tension and thereby decrease vocal hyperfunction, statistical analysis should assess how sensitive RFF was to tracking changes in the pMTD group before and after therapy, as well as how the "original" and "improved" pMTD voices compared to the normal voice (i.e., control) group. In consideration of this study's objectives to explore the clinical potential of RFF to identify the presence of vocal hyperfunction and to quantify change as a consequence of voice therapy, multiple regression analysis and fitting regression lines to the RFF data and calculating their associated slopes was undertaken to examine cycle to cycle variability in RFF across vowel offsets and onsets. Multiple regression analysis was used to examine the effects of cycle, group, and cycle by group interactions on time 1 and time 2 RFF regression line slopes for each of the 3 analyzed tokens: "beautiful" (/ə/-/f/-/əl/), "the shape" (/ə/-/ʃ/-/e/), "shape of" (/e/-/p/-/ə/). The regression analysis showed statistically significant effects that varied across the conditions of token, time, and offset/onset patterns in important ways (Table 3). A statistically significant cycle effect was shown for all but one condition, indicating that both groups displayed RFF slopes that were negative in most conditions. Analysis further demonstrated that in certain conditions, group membership (pMTD or Control) was a significant predictor of differences in RFF cycle values. Group membership significance was shown in onset patterns for all pretherapy conditions and for 2 of the 3 tokens in posttherapy conditions, suggesting that an absolute mean difference in RFF values for onset patterns existed between the groups. No significant group effects existed for any RFF offset patterns, however. Cycle by group interaction 21 Table 3. Multiple Regression Analysis for Cycle and Group Effects. Conditions Cycle Group Cycle*Group T1 O ffset Time 1 β=-0.15, t=-15.93, p=˂.001 β=-0.12, t=-1.44, p=0.169 β=0.07, t=4.99, p=0.001 Time 2 β=-0.32, t=-13.18, p=˂.001 β=-2.36, t=-1.24, p=0.232 β=0.26, t=5.45, p=˂.001 Onset Time 1 β=-0.16, t=-11.97, p=˂.001 β=-0.13, t=-10.15, p=˂.001 β=0.10, t=6.87, p=˂.001 Time 2 β=-0.34, t=-11.38, p=˂.001 β=-1.21, t=-4.65, p=0.003 β=0.15, t=3.54, p=0.003 T2 O ffset Time 1 β=-0.17, t=-7.29, p=˂.001 β=0.19, t=0.92, p=0.373 β=0.01, t=0.37, p=0.714 Time 2 β=-0.35, t=-14.99, p=˂.001 β=-1.90, t=-1.93, p=0.072 β=0.20, t=3.76, p=0.002 Onset Time 1 β=-0.25, t=-12.83, p=˂.001 β=-0.28, t=-8.04, p=˂.001 β=0.09, t=5.26, p=˂.001 Time 2 β=-0.32, t=-13.45, p=˂.001 β=-0.48, t=-2.32, p=0.034 β=0.05, t=1.40, p=0.181 T3 O ffset Time 1 β=-0.07, t=-1.52, p=0.149 β=-0.04, t=-0.09, p=0.930 β=0.02, t=0.30, p=0.765 Time 2 β=-0.22, t=-3.50, p=0.003 β=-0.94, t=-1.06, p=0.306 β=0.12, t=3.31, p=0.005 Onset Time 1 β=-0.06, t=-10.94, p=˂.001 β=-0.17, t=-5.28, p=˂.001 β=0.08, t=4.08, p=0.001 Time 2 β=-0.20, t=-16.52, p=˂.001 β=-0.20, t=-1.83, p=0.086 β=0.14, t=0.79, p=0.444 Significance is bolded. T1=Token 1 (beautiful), T2=Token 2 (the shape), T3=Token 3 (shape of). Time 1 = Pretherapy for the pMTD group, Time 2 = Posttherapy for the pMTD group. effects were statistically significant in 8 of 12 possible conditions, confirming that slopes of the regression lines varied depending upon group membership. Visual representations of RFF mean values for each token by group and therapy time conditions are offered in Figures 3, 4, and 5. Best fit regression lines have been plotted to illustrate overall RFF patterning and highlight differences between groups (pMTD and control) and times (pretherapy and posttherapy). Inspection of Figures of 3, 4, and 5 (combined with the regression analysis above) revealed that the slopes of the regression lines for the pMTD group were less steep as compared to the control group regardless of the token, which is consistent with the hypothesized effects of tension on RFF. That is to say that in healthy phonation, an increase in tension for the devoicing of the consonant translates to an increase in RFF for at least the first several cycles of vibration in the subsequent vowel. In voices affected by muscle tension dysphonia, an already high level of laryngeal tension rendered the increase in RFF following consonant offset less pronounced, and thusly produced a shallower slope. 2􀀕􁔠 Figure 3. Token 1 RFF Mean Values by Group and Time. Token 1 - beautiful (/ə/-/f/-/əl/) 2􀀖􁘠 2􀀗􁜠 Figure 5. Token 3 RFF Mean Values by Group and Time. 2􀀘􁠠 RFF onsets as compared to offsets also appeared to be more sensitive to not only group differences, but also treatment effects. That is, the nonsignificant beta coefficients for the cycle by group interaction observed for tokens 2 and 3 RFF onsets (for time 2) suggest that the slopes of the regression lines for the posttreatment pMTD and controls are not significantly different. This pattern of nonsignificance in RFF onset supports that the posttreatment slope for the pMTD group was similar to the controls, and was consistent with a putative reduction in vocal hyperfunction following treatment. (See Figures 3, 4, and 5.) To better understand the influence of group, time, and age on the slope of RFF values, a slope coefficient was calculated for each subject for both offset and onset RFF patterns for each of the 3 analyzed tokens. Table 4 summarizes the results of a multiple regression analysis examining the possible influence of the variables group membership (pMTD vs. Control), time (time 1 vs. time 2), age, and their respective interactions on slope as the dependent variable. The analysis revealed that slope was not influenced by age, nor were there significant group by time or group by age interactions (all beta coefficients = 0.00, p>.05). However, significant main effects were observed for group membership, as well as for time condition in RFF onset patterns. These findings from the regression analysis confirmed that the slopes of RFF onset consistently varied as a function of group membership (pMTD vs. Controls) and time (time 1 vs. time 2). The results also suggested that the slopes of RFF offset appeared to be less consistently influenced by group and time as evidenced by variable significance patterns. Thus, RFF onsets appeared to be a more stable (predictable) environment to reveal group and time effects. 26 Table 4. Multiple Regression Analysis for Group, Time, and Age Effects. Conditions Group Time Age Token 1 Offset β=0.12, t=3.32, p=0.001 β=-0.01, t=-0.04, p=0.665 t=-1.69, p=0.09 Onset β=0.24, t=3.88, p=0.001 β=0.11, t=2.65, p=0.009 t=-1.13, p=0.26 Token 2 Offset β=0.06, t=1.6, p=0.112 β=0.02, t=0.85, p=0.395 t=-0.47, p=0.64 Onset β=0.12, t=2.34, p=0.020 β=0.10, t=3.06, p=0.003 t=0.07, p=0.95 Token 3 Offset β=0.10, t=2.12, p=0.035 β=-0.06, t=-1.98, p=0.049 t=-1.75, p=0.08 Onset β=0.11, t=2.19, p=0.030 β=0.07, t=2.07, p=0.040 t=-1.46, p=0.15 Significance is bolded Token 1 (beautiful), Token 2 (the shape), Token 3 (shape of). RFF and Dysphonia Severity Another focus of this study was to determine the extent to which RFF related to listener-perceptual ratings of dysphonia severity. The initial research question in this study analyzed RFF slope, or the overall pattern created by all 10 cycles both preceding and following a voiceless consonant. In contrast, several previous studies (Stepp et al., 2010; 2011; 2012) have reported that aggregated RFF means for cycle 10 offset and cycle 1 onset, the points furthest from the normalization cycles and therefore the values most likely to differentiate levels of hyperfunction, to be correlates of dysphonia severity. This type of single RFF cycle analysis was used in the current study to examine the secondary question of whether a correlation exists between RFF and listener-perceptual ratings of dysphonia severity. A summary of cycle 10 offset and cycle 1 onset means, ranges, and standard deviations for each of the 3 tokens analyzed in this study is provided in Table 5. The RFF means for offset cycle 10 and onset cycle 1 across tokens are presented here for comparison purposes to those studies, but because of token-to-token variability, were not used for statistical analysis. The mean RFF values for all 10 offset and onset cycles for each of the 3 analyzed tokens are displayed in Appendix A. While mean semitones (ST) for all measures were somewhat different than those found in previous studies (Stepp et 27 Table 5. RFF Means, Standard Deviations, Ranges, and Numbers by Token. Pretherapy Posttherapy Conditions N Mean SD Min Max N Mean SD Min Max T1, Offset 10 pMTD 57 -0.69 1.75 -4.77 3.49 94 -0.37 1.31 -3.68 3.19 Control 18 -1.35 1.53 -3.85 1.93 18 -1.27 1.73 -4.86 1.41 T1, Onset 1 pMTD 40 0.87 2.65 -4.02 7.62 54 1.38 2.08 -5.42 5.55 Control 12 2.98 1.37 0.25 4.76 12 2.54 2.00 -1.51 5.17 T2, Offset 10 pMTD 27 -1.44 1.79 -4.30 3.39 45 -1.32 1.47 -4.09 3.39 Control 17 -1.23 1.74 -3.77 2.96 15 -1.87 1.49 -4.97 0.72 T2, Onset 1 pMTD 70 1.31 2.40 -4.97 7.91 108 2.52 2.24 -3.60 7.40 Control 19 2.43 1.74 0.05 5.94 20 2.67 1.81 -0.23 6.07 T3, Offset 10 pMTD 71 -0.12 2.21 -9.05 4.93 92 0.35 1.40 -2.51 4.69 Control 18 -0.18 1.65 -3.37 2.61 18 -0.54 1.22 -2.93 2.83 T3, Onset 1 pMTD 49 0.90 1.99 -2.81 5.01 73 1.55 2.03 -3.82 8.17 Control 16 1.58 1.94 -0.83 5.62 17 1.60 1.60 -0.52 4.50 All, Offset 10 pMTD 83 -0.52 1.93 -9.05 4.93 110 -0.18 1.06 -2.90 3.32 Control 20 -0.95 0.89 -2.18 1.16 20 -1.03 1.13 -3.28 1.12 All, Onset 1 pMTD 84 1.10 2.10 -4.97 7.91 110 1.98 1.62 -3.55 6.33 Control 20 2.15 1.24 -0.39 4.38 20 2.34 1.13 0.55 4.29 T1=Token 1 (beautiful), T2=Token 2 (the shape), T3=Token 3 (shape of). Offset 10=RFF Offset Cycle 10, Onset 1=RFF Onset Cycle 1. al., 2010, 2011, 2012), overall trends were similar with the exception of the RFF offset cycle 10 values for the control group (which were substantially lower). Using these specific RFF values, we explored the extent to which RFF was predictive of severity rating, or more specifically, whether specific RFF values could index increases in the severity of vocal hyperfunction as determined by listener dysphonia ratings on the 0-100 scale. Previous research has established almost perfect correlations between perceptual ratings of "vocal effort" (VE) and overall dysphonia severity, and confirmed that VE accounted for 97% of the variance in listener ratings of overall severity (Eadie & Stepp, 2013). Therefore, VE and overall dysphonia severity are almost synonymous. In this regard, Eadie and Stepp (2013) examined correlations between vocal effort (VE) and the specific RFF values associated with offset cycle 10 28 and onset cycle 1. As mentioned, these 2 RFF values were used because it was felt that those two measures were "most likely to differentiate among levels of vocal hyperfunction, given the patterns seen previously in this population (Stepp et al., 2010, 2011) and because they were the points furthest from those used for normalization." Likewise in this study, Pearson product-moment correlations were used to examine possible correlations between dysphonia severity and the RFF offset 10 and onset 1 cycle values for each of the three evaluated tokens (Table 6). Given that the severity ratings in the control group indicated normal voice quality, and by extension, no vocal effort or hyperfunction, analysis was only performed for the pMTD group. Although individuals with pMTD were "improved" following treatment, they were not uniformly judged as "normal" by listener ratings. For this reason, correlation analysis was completed for both pretherapy and posttherapy conditions. A significant correlation was observed between RFF onset 1 cycle and severity rating for tokens 1 and 2. There was no significant Table 6. Correlations of RFF Value and Severity. Conditions Pearson Product-Moment Correlations Number Token 1 Offset 10 Time 1 r=0.17, p=0.217 57 Time 2 r=0.03, p=0.755 94 Onset 1 Time 1 r=-0.32, p=0.042 40 Time 2 r=-0.32, p=0.018 54 Token 2 Offset 10 Time 1 r=0.11, p=0.582 27 Time 2 r=0.15, p=0.316 45 Onset 1 Time 1 r=-0.38, p=0.001 70 Time 2 r=-0.28, p=0.004 108 Token 3 Offset 10 Time 1 r=-0.03, p=0.795 71 Time 2 r=-0.03, p=0.774 92 Onset 1 Time 1 r=-0.16, p=0.260 49 Time 2 r=-0.16, p=0.17 73 Significance is bolded. Token 1 (beautiful), Token 2 (the shape), Token 3 (shape of). Time 1 = Pretherapy for the pMTD group, Time 2 = Posttherapy for the pMTD group. Offset 10=RFF Offset Cycle 10, Onset 1=RFF Onset Cycle 1. 29 relationship between RFF offset 10 cycle and severity rating. Another regression analysis considered the influences of therapy condition (time) and the specific RFF cycles of offset 10 and onset 1 on dysphonia severity rating. Analysis was performed first for just the pMTD and subsequently for the groups combined. Results were similar and so only the combined analysis has been reported (Table 7). This analysis was consistent with other findings from this study, confirming that RFF onset may be sensitive to predicting the presence and degree of vocal hyperfunction (as indexed by overall dysphonia severity), both before and after therapy. Missing/Unanalyzable RFF Data Previous studies have reported that glottalization, inability to identify cycle boundaries, (due to insufficient periodicity), a lack of a sufficient "steady-state," vowel brevity, or a grammatical pause taken by the speaker might render it impossible to capture the requisite 10 cycles of vibration or to otherwise reliably calculate RFF for a specific token (Stepp et al., 2011). In our study, it became apparent during RFF calculation that numerous RFF patterns could not be calculated and similarly would need to be excluded from analysis. This led to missing data for the 3 tokens examined. Table Table 7. Regression Analysis for RFF Cycle and Time Effects on Severity. Conditions RFF Cycle Time Token 1 Offset 10 β = 2 . 8 2, t=2.36, p=0.019 β=35.89, t=9.54, p˂0.001 Onset 1 β=-4.57, t=-4.71, p˂0.001 β=37.49, t=8.33, p˂0.001 Token 2 Offset 10 β = 1 .14, t=0.67, p=0.506 β=25.32, t=4.63, p˂0.001 Onset 1 β=-3.51, t=-4.91, p˂0.001 β=30.75, t=9.30, p˂0.001 Token 3 Offset 10 β = 0 .02, t=0.02, p=0.99 β=35.82, t=10.02, p˂0.001 Onset 1 β=-2.03, t=-2.03, p=0.044 β=32.20, t=8.11, p˂0.001 Significance is bolded. Token 1 (beautiful), Token 2 (the shape), Token 3 (shape of). Offset 10=RFF Offset Cycle 10, Onset 1=RFF Onset Cycle 1. 30 8 presents a summary of RFF patterns that were excluded from the analysis set, coded as to why the exclusions were made. The mean number of tokens analyzed for each individual subject was 1.87 out of a possible 3 and Table 9 provides additional information about the average number of usable tokens across group and therapy time conditions. Adjusted odds ratio (OR) estimates (adjusted for all other variables in the model) were used to assess factors influencing the missing data among the 3 tokens used for RFF analysis. ORs can range from 0 to infinity. The OR is statistically significantly different than 1.0 at the 0.05 level of significance if the 95% confidence interval (CI) does not include 1.0. An OR of less than 1 indicates a negative association, while an OR of 1 indicates no association, and an OR of more than 1 indicates a positive association between two variables. If both the lower and upper limits of the CI are less than 1, there is a significant negative association, whereas if both the lower and upper limits of the CI are greater than 1, there is a significant positive association. Factors shown to be Table 8. Number of Excluded RFF Patterns (out of 262 Possible). RFF Offset RFF Onset Token 1 FC: 19 NC: 25 SS: 31 Total: 75 28.63% FC: 3 NC: 96 SS: 45 Total: 144 54.96% Token 2 FC: 70 NC: 38 SS: 49 Total: 157 59.92% FC: 3 NC: 22 SS: 20 Total: 45 17.18% Token 3 FC: 5 NC: 35 SS: 24 Total: 64 24.43% FC: 2 NC: 76 SS: 31 Total: 109 41.60% FC - fewer than 10 cycles displayed by Praat; NC - no cycles displayed by Praat; SS - sample was not at ‘steady-state'. Table 9. Mean Number of Analyzable Tokens. Group Time 1 Time 2 pMTD Pretherapy: 1.41 Posttherapy: 2.10 Combined: 1.75 Controls 2.50 2.50 2.50 Combined group and time Overall Mean: 1.87 31 significant were token, group, and severity rating. It was revealed that the odds of missing data were significantly greater for tokens 1 and 2 (as compared with token 3) for pMTD subjects, as well as among those subjects presenting with greater listener dysphonia severity scores (Table 10). It was originally suggested that a "severity threshold" might exist, beyond which RFF calculation proved too unreliable to be clinically viable. Missing data, or the impossibility of calculating RFF, might be one way to represent such a severity threshold. Adjusted OR estimates demonstrated that token, group, and severity ratings all contributed to the likelihood of missing data points, so Table 11 was developed in an attempt to illuminate where a hypothetical severity threshold might exist in relation to group, therapy time, and token conditions. The pMTD group severity ratings were rank-ordered and then divided into quartiles for pretherapy and posttherapy conditions. Dysphonia severity ratings and associated missing data points were calculated by quartile. Missing data points existed in the control group as well as in the pMTD patient sample, but because the control group contained perceptually normal voices, it was not necessary to divide the group based on severity. This table illustrates a substantial and variable attenuation of data across severity quartiles, time (pre- or posttherapy) and phonetic context (token). Missing data were consistently higher in the pMTD group as compared with controls. Numbers and Table 10. Odds Ratio Estimates.* Effect Point Estimate 95% Wald Confidence Limits Token 1 vs 3 1.569 1.201 - 2.050 Token 2 vs 3 1.348 1.031 - 1.764 Group 1.748 1.190 - 2.569 Severity 1.017 1.014 - 1.021 *Adjusted for all variables included in the model. Token 1 (beautiful), Token 2 (the shape), Token 3 (shape of). 32 Table 11. Missing Data by Severity Quartile. Severity n= pMTD, Time 1 (Pretherapy) T1OFF T1ON T2OFF T2ON T3OFF T3ON <46 24.3% 25.93% 51.85% 66.67% 7.41% 7.41% 29.63% 46-<75 26.1% 44.83% 65.52% 72.41% 31.03% 31.03% 44.83% 75-<95 24.3% 40.74% 70.37% 74.07% 33.33% 33.33% 74.07% 95-100 25.2% 40.38% 67.86% 85.71% 75% 67.86% 75% No. Missing Overall 52 71 83 41 39 62 % Missing Overall 46.8% 64.0% 74.8% 36.9% 35.1% 55.9% pMTD, Time 1(Posttherapy) T1OFF T1ON T2OFF T2ON T3OFF T3ON <3.9 25.23% 14.29% 39.29% 57.14% 0% 14.29% 32.14% 3.9-<8.3 25.23% 17.86% 60.71% 53.57% 3.57% 21.43% 28.57% 8.3-<18 24.33% 11.11% 51.85% 59.26% 3.7% 11.11% 40.74% 18-<73 25.23% 10.71% 53.57% 53.57% 3.57% 21.43% 35.71% No. Missing Overall 15 57 62 3 19 38 % Missing Overall 13.5% 51.4% 55.9% 2.7% 17.1% 34.2% Control, Time 1 T1OFF T1ON T2OFF T2ON T3OFF T3ON <10 100% 10.0% 40.0% 15.0% 5.0% 10.0% 20.0% No. Missing Overall 2 8 3 1 2 4 % Missing Overall 10.0% 40.0% 15.0% 5.0% 10.0% 20.0% Control, Time 2 T1OFF T1ON T2OFF T2ON T3OFF T3ON <12 100% 10.0% 40.0% 25.0% 0.0% 10.0% 15.0% No. Missing Overall 2 8 5 0 2 3 % Missing Overall 10.0% 40.0% 25.0% 0.0% 10.0% 15.0% Token 1 (beautiful), Token 2 (the shape), Token 3 (shape of). OFF=RFF Offset, ON=RFF Onset. Time 1 = Pretherapy for the pMTD group, Time 2 = Posttherapy for the pMTD group. percentages of missing data also increased steadily across the pretherapy condition for the pMTD group, potentially to such a degree that the subjects in the more severely disordered quartiles were under-represented in the data used for statistical analysis. For instance, because quartiles were divided so that each had an approximately equivalent number of subjects, 27 individuals had severity ratings of 95-100 prior to therapy. Missing data ranged from around 40% to over 85% depending on token and RFF 33 offset or onset condition. This means that depending on RFF pattern, anywhere from 4 to 16 individuals contributed data for statistical analysis. In contrast, anywhere from 9 to 25 individuals had analyzable data in the first quartile before therapy. This potentially led to an under-representation of voices characterized by more severe vocal hyperfunction in the final RFF analysis. While missing data were most prevalent in the pMTD group, the control group was not immune to data exclusion. Great variability among tokens and RFF offset/onset conditions suggested that phonetic context was an important influence on missing data, and dysphonia severity alone does not account for all excluded or incalculable RFF values. DISCUSSION RFF Onset Values across Group and Time Conditions In the field of voice disorders, vocal hyperfunction related to excess laryngeal musculoskeletal tension is a recurrent theme, and is featured prominently in the disorder pMTD. This study sought to examine the clinical value of RFF as means to identify and track changes in vocal hyperfunction before and following a course of manual circumlaryngeal therapy for patients with pMTD. The regression and associated analyses of RFF slopes confirmed that RFF onset has the potential to serve as a clinical index of hyperfunction. RFF onset slopes differed significantly based upon group membership (disordered or normal) and between pre- and posttherapy time conditions. Significant between group differences confirmed that RFF onset slope may be a viable means to distinguish between voices influenced by vocal hyperfunction and those free from abnormal amounts of musculoskeletal tension, while significant time effects represent RFF's sensitivity to tracking change across the course of voice therapy. Both group and time effects were consistently significant in RFF onset slopes, but not for RFF offset slopes, suggesting that RFF offset may be a less sensitive/predictable context for measuring hyperfunction. These conclusions were further supported by the exploration of cycle, group, and cycle by group interaction effects. Recall that group membership was significant in the onset condition for all pretherapy tokens (Table 3). Similarly, cycle by group interaction 35 effects were significant for the same pretherapy onset slopes. In posttherapy conditions, a group effect was present in 2 of 3 tokens and an interaction effect was significant for 1 of 3 tokens. Because the overarching goal of voice therapy was to eliminate tension, one could surmise that fewer adverse effects were shown following therapy because tension, or the presence of vocal hyperfunction, was reduced sufficiently for pMTD voices to approximate normal phonation, at least in some phonetic contexts. These analyses suggest that RFF was a sufficiently sensitive acoustic measure to identify the presence of vocal hyperfunction in the pMTD group before therapy in RFF onset slopes. A significant "cycle" main effect in the same analysis was indicative of an aggregated downward slope of all regression lines, indicating that both groups displayed slopes that were negative. While downward slopes were anticipated for RFF onset patterns, they represent something of a departure from previous research that found that normal voices demonstrated a more level pattern for RFF offset, or in effect, no significant cycle to cycle variability (Goberman & Blomgren, 2008; Stepp et al., 2010). In fact, the mean RFF values for the control group in this study were lower than the mean RFF values for the pMTD cases in the offset condition, a direct contrast to findings from previous research showing that control RFF offset values were consistently higher than those displayed by the disordered population (Stepp et al., 2010; 2011; 2012). The RFF onset condition for all 3 tokens, however, displayed RFF values consistent with previous studies. Based on this study's findings, RFF onset patterns may be more reliable than offset patterns when analyzing voices suspected of vocal hyperfunction. Following therapy, onset RFF values for the pMTD group were shown to more closely approximate those found in the normal group (Figures 3, 4, 5). Indeed, slope was 36 not shown to be significantly different following therapy for tokens 2 and 3, and while still lower than the control group, RFF values were closer to normal than prior to treatment. Significant "group" effects in the completed analysis captured magnitude differences in mean RFF across all cycles. This showed that even though the slopes of both groups were parallel following therapy, their absolute RFF values were not the same, with the pMTD being lower than the normal group. Therefore, while voice therapy presumably lessened muscle tension substantially, it may not have eliminated vocal hyperfunction for some or all patients in the pMTD group. This was consistent with post-treatment listener-perceptual severity ratings, which ranged largely from 0-40 on the 100 point scale. While posttherapy scores were within or closer to the normal/control range for the pMTD group, auditory-perceptual symptoms of dysphonia persisted in some cases. In other words, RFF onset patterns showed sensitivity to differences in the absolute magnitude of RFF values between groups in both pretherapy and posttherapy conditions, whereas offset patterns did not reveal significant differences and were not sensitive to change resulting from therapy. This finding demonstrates the clinical potential of RFF onset patterns to quantify change in what is assumed to be vocal hyperfunction (excess tension) across the course of voice therapy. It is important to remember that the controls in this study did not behave like the controls from previous research in terms of RFF offset patterning. Previous research has found significant differences in RFF offset patterns between disordered and control groups (Stepp et al., 2012), albeit by way of different statistical analyses and slightly varied patient populations. Had the controls from this study behaved more similarly to those examined previously, significant differences may have also been described for RFF 37 offset patterns in this study. A more extensive examination of RFF offset and onset patterns in a large sample of normal voices would be advisable in order to begin the process of establishing RFF norms and the optimal conditions for RFF calculation. Correlations between RFF and Listener Perceptual Ratings Voices presenting with severe dysphonia are not difficult to qualitatively assess with auditory-perceptual measures as being impaired. That is to say, a voice recording from a subject in this study with a high severity rating (e.g., 60-100) would be identified as disordered even by the most inexperienced listener. The degree of impairment, however, especially in "borderline" cases, may represent a challenge in consistency when subjective measures alone are used for clinical assessment. This study's exploration into possible correlations between listener severity ratings and RFF offset cycle 10 and onset cycle 1 values hinted at a clinical potential for objectively quantifying varying degrees of dysphonia severity resulting from vocal hyperfunction. Pearson product-moment correlations showed a relationship between severity ratings and RFF onset cycle 1 values for tokens 1 and 2 in both pre- and posttherapy conditions (Table 6). No correlations were found for RFF offset cycle 10 values for any of the 3 examined tokens. These findings were similar to those from previous research that found a correlation between vocal effort ratings and RFF onset cycle 1 values (Eadie & Stepp, 2013). Such findings indicate that as RFF onset cycle 1 values lower, severity ratings tend to increase. The regression analysis exploring cycle and time effects on severity confirmed that RFF onset cycle 1 consistently influenced severity ratings across all 3 tokens (Table 7). Looking to the future, RFF onset cycle 1 could potentially be used as an index of 38 severity, ostensibly quantifying the degree of vocal hyperfunction in disordered voices. Furthermore, a divergence from established norms could objectively indicate the presence of dysphonia or vocal hyperfunction, even if subjectively the voice was classified as within normal limits. This could assist clinicians in best determining when to begin and end treatment for patients presenting with perceptually normal or near-normal voice quality, especially following voice therapy. One important caveat is that the analysis from this study is only representative of a subset of patients and controls due to high levels of missing data, especially among the more severely disordered voices. Therefore, onset RFF may only be able to quantify increasing levels of dysphonia for mild to moderate levels of hyperfunction, and not across the entire severity spectrum. Dysphonia Severity and RFF Calculations If one accepts severity of dysphonia as an index of vocal hyperfunction (i.e., excess tension in voice production), then this investigation explored the entire spectrum of vocal hyperfunction, from voices produced with no excess tension to those maximally affected (i.e., dysphonia rating range of 0-100). Severity ratings pretherapy for the pMTD group averaged 67.9 (SD=27.8). The average VE and OS scores for disordered voices in one previous study were 25.8 and 23.5, respectively (Stepp et al., 2012), confirming that we examined patients with more severe symptoms of vocal hyperfunction than prior research. Perhaps as a consequence, dysphonia severity ratings and missing data seemed to be linked in this study. This was most noticeable for the pMTD group in the pretherapy condition (Table 11). As severity of dysphonia increased across the four quartiles, the number of usable/measureable RFF data consistently decreased. This observation would seem to support the notion of a severity threshold, beyond which the 39 viability of RFF as a clinical measure diminishes substantially. However, adjusted OR estimates showed that not only severity, but also group and token were contributing factors to missing data (Table 10). Group membership implies the presence of vocal hyperfunction and is inextricably linked to severity. Clearly the pMTD group had higher severity scores than the controls both before and after therapy, and was associated with greater odds of missing data. Token (i.e., phonetic context) is the most plausible explanation for missing data in the control group given that severity was consistently within normal limits and subjects were not affected by hyperfunction. In contrast, a combination of token and severity likely influenced missing data in the pMTD group, both before and after therapy. For instance, the data revealed that beyond a severity score of 45 (on the 0 to 100 severity rating scale) for the pMTD group in the pretherapy condition, missing data became increasingly frequent for all tokens for both RFF offset and onset patterns. However, a similar observation of a steadily increasing missing data cannot be made in the posttherapy condition. One would anticipate that if a severity threshold truly existed, the fourth quartile in the posttherapy condition (ranging 18-73 points) would demonstrate a higher loss of data than the three other quartiles (ranging 0-17.9 points) with severity ratings in the normal to mild ranges only. The numbers of missing data points across all posttherapy quartiles were instead relatively constant. This suggests that rather than severity, VCV token was a seemingly stronger influence on missing data posttherapy. In other words, RFF data were unusable not because of the aperiodicity or aphonia associated with increasingly dysphonic voices, but instead because of inherent challenges presented by each VCV token's specific phonetic characteristics. Severity still affected 40 missing data in the posttherapy condition though, because despite being steady across quartiles, percentages of missing data for the pMTD group following therapy were higher than those in the control group. To summarize, a combination of token and severity contributed to missing data in both treatment conditions, although group and severity seemingly played a far greater role in data loss pretherapy. Given this apparent combination of influencing factors, it is challenging to reliably determine where a hypothetical severity threshold might lie without controlling for VCV token (i.e., phonetic context), making continued research into the optimal phonetic contexts for RFF calculation advisable. Tokens should not be investigated only in mild to moderately impaired voices, as this risks introducing spectrum bias and thereby diminishing the clinical viability of RFF. There may be phonetic contexts that allow for more reliable RFF calculation than others across the severity spectrum, but until standardized VCV tokens are established, it will be challenging to conclusively determine RFF's applicability to more severely disordered voices. Causes of Missing/Unanalyzable Data A large number of RFF offset and onset patterns were eliminated from analysis because of aperiodicity, glottalization, vowel brevity, or lack of steady state phonation (Table 11). As described, numbers of missing data were significantly greater in the pMTD group, especially prior to therapy when dysphonia severity ratings were proportionally highest. However, it is important to re-iterate that even when posttherapy pMTD subjects had "normal" ratings suggesting no dysphonia, there were still more missing data than in the control group. This might indicate that despite sounding subjectively normal, an underlying (subclinical) tension or hyperfunction persisted in the 41 pMTD group, rendering them less likely to present at "steady-state" phonation or with similar vowel durations as compared with controls. The ability to consistently calculate RFF might then be a distinguishing factor between truly normal voices and those with borderline or mild dysphonia otherwise undetectable from auditory-perceptual evaluation alone. OR estimates demonstrated that as severity of vocal hyperfunction increased, the chance of successfully calculating RFF for all offset and onset patterns of any given number of tokens decreased. For instance, the average number of tokens calculable for all subjects across all conditions in this study was 1.87. This is markedly fewer than the 2.5 average analyzable in a similar study (Stepp et al., 2012). The mean for controls, however, was 2.5 consistently across time, while the mean for the pMTD group increased from 1.41 in the pretherapy condition to 2.10 in the posttherapy condition. All other factors being constant, it would seem that the reduction in severity accomplished through voice therapy directly led to an increase in analyzable data. One would assume this was because the pMTD group had a reduction in hyperfunction, producing more steady state, periodic signals in which RFF was calculable. The mean number of analyzable tokens approached but did not equal the control group, providing evidence that RFF was sensitive to the presence of and changes in vocal hyperfunction. However, the calculation of RFF requires analysis of a time-based waveform (and a quasiperiodic acoustic signal) and this potential limitation is highlighted by the fact that as severity of dysphonia increased (i.e., presumably more type 2 and 3 signals and fewer periodic signals), mean number of analyzable data decreased. This dependence on periodicity represents a critically important limitation to using RFF as an objective assessment 42 measure with severely disordered voices (affected strongly by vocal hyperfunction), which characterizes many patients with pMTD. Severity of dysphonia alone did not account for all missing data (Table 11). Phonetic context (i.e., token) was the most likely explanation for the missing data in the control group, most typically because of vowel brevity, but also possibly due to speaker pauses or a lack of "steady-state" phonation. It is important to consider that phonetic context does not only mean the components of the VCV token, but also implies the position of that token in a word, phrase, sentence, and/or paragraph. For instance, token 1 ("beautiful") is followed by the word "colors," which concludes a sentence. Many speakers used the /k/ in "colors" to truncate the vocalic /l/ of beautiful, possibly leading to missing data as a result of too few cycles of vibration necessary to calculate RFF. In fact, each token selected for analysis from The Rainbow Passage presented certain challenges to RFF calculation. Despite the lengthy fricative consonant in token 1 (/f/), a consonant demonstrated in previous research to be optimal for RFF calculation (Lien, et al., 2014), both vowels were prone to brevity in the reading sample. Vowel brevity was also a factor for the offset of token 2 ("the") and the onset of token 3 ("of"), again decreasing the potential for phonation to reach "steady-state" in running speech, as well as allowing for fewer than the 10 necessary cycles of vibration. The fewest number of missing data were seen for the onset of token 2 and the offset of token 3 (Table 8). The /e/ in the word "shape" was the vowel used in both these contexts, and it is worth noting that it was both longer in duration and the only accented or non-schwa vowel used for analysis. Lien, Gattuccio, and Stepp (2013) described a lack of knowledge regarding the effect and potential differences that various phonemes in VCV tokens might have on 43 RFF, and while their study began this exploration, future research is necessary to determine optimal and standardized tokens for use in a clinical voice assessment protocol. Previous studies have had somewhat more consistent results using isolated tokens in preference to running speech contexts for RFF extraction (Lien, 2014). While running speech may arguably be a more ecologically valid way to examine a client's typical speaking patterns, VCV stimuli's contribution to the number of missing data in this study highlights one challenge inherent in analysis of running speech. One alternative to using isolated phoneme patterns (e.g., /isi/, /apa/) would be to develop phonetically-optimized sentences for RFF calculation, perhaps similar to the voiceless-consonant laden sentence used clinically for assessment of patients symptomatic of spasmodic dysphonia (Eadie & Stepp, 2013). Continued research into the influence of phonetic context on RFF is necessary to identify ideal token and sentence stimuli. Along these same lines, it was notable that unlike previous studies, the tokens analyzed in this research displayed significant cycle by token interaction effects (Table 2). This made averaging data across tokens inappropriate. Determining precisely why each token behaved uniquely can only be speculated, but it seems probable that the phonetic characteristics of running speech were influential on vowel length and phonation stability, especially when schwas were included in the VCV token. Another possibility is that varying numbers of missing data by token offset and onset may have skewed the range of dysphonia severity represented in the analyzed data, resulting in token behavior variance. Future research may help to illuminate underlying reasons why RFF may vary by token, an area of exploration necessary before a standardized protocol 44 for RFF calculation can be introduced clinically. Automation of RFF Calculation Manually calculating RFF values for 3 or more VCV tokens takes a significant amount of time (Lien & Stepp, 2013; Stepp et al., 2010). Lien and Stepp (2013) recognized this as a limitation to future research and clinical application/use considering that while "the mechanism underlying the observed RFF has been hypothesized as the interplay of tension, aerodynamics, and vocal fold kinematics, the contribution of each mechanism is not clear. Elucidation of the physiological mechanisms that result in the characteristic RFF in healthy individuals and the changes that occur to these mechanisms with voice disorders are essential for clinical validation of RFF" (p. 1). To this end, they examined an automated method of estimating RFF with the goal of expediting the analysis of multiple tokens and making research of larger sample sizes more feasible. Their proposed algorithm produced estimates similar to those calculated manually. However, the automation could not be used reliably in the analysis of connected, or coarticulated, speech. Automating the RFF calculation process would both increase the measure's objective consistency, as well as the potential for analysis of larger subject samples. Objective consistency is mentioned because when choosing offset and onset cycles using Praat in this study, an element of the investigators' "clinician's best judgment" or subjectivity was required. The recordings included numerous type 2 and 3 signals in the pMTD group that are known to be notoriously difficult to objectively analyze in the time domain (Titze, 1995) as they lack regular periodicity (i.e., cycle boundaries are difficult to establish). As a consequence, Praat may have inaccurately 45 identified cycle boundaries in some instances. Additionally, the primary investigator in this research selected a zoom level or 0.3-0.5 seconds for pulse extraction to identify the 10 cycles of vibration preceding and following voiceless consonant stimuli. A zoom level was not specified in any prior studies, but zoom level as well as horizontal movement (scrolling) on the waveform display in Praat frequently influenced pulse placement by adding or subtracting one pulse on either end of the voiceless consonant, and therefore changed vibratory cycle identification. The development of precise guidelines specific to RFF calculation could help minimize subjectivity and improve interjudge reliability until such a time as this measure may be automated. CONCLUSION This study lent support to the assertion that RFF has clinical potential as an objective, noninvasive measure of vocal hyperfunction. It is encouraging that while offset RFF patterns examined for this study, for whatever reason, were not sensitive to group and therapy-condition differences, RFF onset patterns were. Specifically, RFF onset slopes and cycle 1 values before and following therapy were shown to be sensitive to the presence of and changes in severity of vocal hyperfunction. Onset RFF patterns may therefore be clinically valid and valuable for quantifying change in patients experiencing vocal hyperfunction across the course of voice therapy, as well as perhaps identifying the presence and degree of vocal hyperfunction in patients across at least a portion of the severity spectrum. However, phonetic context (token) and severity level of dysphonia were shown to significantly influence the performance, validity, and clinical utility of RFF. Identifying the optimal phonetic contexts for utilizing the RFF measure is something that has only recently been explored (Lien et. al., 2013), although it is likely that running speech will be less ideal than isolated VCV tokens, especially when considering that a proposed algorithm to automate RFF calculation has only been successful with isolated tokens. An investigation into the most optimal VCV tokens for RFF analysis to minimize data exclusion and variability across tokens while maximizing the range of assessable clients regardless of vocal hyperfunction severity seems warranted. This would allow for the 47 development of standardized voice assessment protocols, a necessary step for RFF's inclusion in a clinical setting. Further exploration of the potential relationship between RFF and the gold standard of auditory-perceptual ratings may also demonstrate clinical value. Overall, RFF onset values only were consistent with previous studies and this study's original hypotheses. However, it is critically important to recognize that our reported results represent only a subset of the total number of subjects evaluated in both the pMTD and control groups due to a large number of missing or necessarily excluded data. Due to the unique influence of group and severity on missing data, it is unlikely that the results are representative of RFF patterns in dysphonic voices on the more severe end of the severity spectrum. The accurate calculation of RFF (whether manually or automated) is dependent upon a quasiperiodic voice signal (where the fundament period is easily extracted), thus this dependence represents a significant shortcoming. In light of this limitation, further research will need to be undertaken to better understand the role of RFF in clinical assessment and as a treatment outcomes measure. APPENDIX Token 1 (beautiful) Time 1 (Pretherapy) Time 2 (Posttherapy) Cycle N pMTD N Controls N pMTD N Controls Offset 10 59 0.00 18 0.00 96 0.00 18 0.00 Offset 9 57 -0.08 18 -0.08 94 -0.06 18 0.01 Offset 8 57 -0.05 18 -0.15 94 -0.05 18 -0.16 Offset 7 57 -0.21 18 -0.32 94 -0.09 18 -0.41 Offset 6 57 -0.26 18 -0.48 94 -0.21 18 -0.65 Offset 5 57 -0.27 18 -0.63 94 -0.41 18 -0.86 Offset 4 57 -0.45 18 -0.77 94 -0.49 18 -1.06 Offset 3 57 -0.74 18 -0.83 94 -0.55 18 -1.19 Offset 2 57 -0.58 18 -1.23 94 -0.62 18 -1.16 Offset 1 57 -0.69 18 -1.35 94 -0.37 18 -1.27 Onset 1 40 0.87 12 2.98 54 1.38 12 2.54 Onset 2 40 0.13 12 1.84 54 1.61 12 3.02 Onset 3 40 0.20 12 2.24 54 1.49 12 2.10 Onset 4 40 0.45 12 1.76 54 0.99 12 1.28 Onset 5 40 0.32 12 1.09 54 0.81 12 1.02 Onset 6 40 0.26 12 1.11 54 0.52 12 0.74 Onset 7 40 0.07 12 0.48 54 0.40 12 0.52 Onset 8 40 0.11 12 0.27 54 0.25 12 0.23 Onset 9 40 0.08 12 0.16 54 0.12 12 0.05 Onset 10 40 0.00 12 0.00 57 0.00 13 0.00 49 Token 2 (the shape) Time 1 (Pretherapy) Time 2 (Posttherapy) Cycle N pMTD N Controls N pMTD N Controls Offset 10 28 0.00 17 0.00 49 0.00 15 0.00 Offset 9 27 -0.07 17 -0.06 46 -0.13 15 -0.03 Offset 8 27 -0.12 17 -0.24 46 -0.29 15 -0.18 Offset 7 27 0.20 17 -0.33 46 -0.38 15 -0.29 Offset 6 27 -0.02 17 -0.48 46 -0.47 15 -0.49 Offset 5 27 -0.29 17 -0.70 46 -0.55 15 -0.83 Offset 4 27 -0.46 17 -1.01 46 -0.86 15 -1.22 Offset 3 27 -0.93 17 -1.27 46 -1.11 15 -1.65 Offset 2 27 -1.07 17 -1.47 46 -1.37 15 -1.92 Offset 1 27 -1.44 17 -1.23 45 -1.32 15 -1.87 Onset 1 70 1.31 19 2.43 108 2.52 20 2.67 Onset 2 70 1.14 19 2.92 108 2.04 20 2.36 Onset 3 70 0.64 19 2.70 108 1.52 20 1.95 Onset 4 70 0.73 19 1.77 108 1.12 20 1.82 Onset 5 70 0.59 19 1.28 108 0.78 20 1.21 Onset 6 70 0.33 19 0.99 108 0.60 20 0.70 Onset 7 70 0.27 19 0.58 108 0.40 20 0.48 Onset 8 70 0.11 19 0.27 108 0.25 20 0.25 Onset 9 70 0.00 19 0.17 108 0.11 20 0.14 Onset 10 73 0.00 20 0.00 108 0.00 20 0.00 50 Token 3 (shape of) Time 1 (Pretherapy) Time 2 (Posttherapy) Cycle N pMTD N Controls N pMTD N Controls Offset 10 72 0.00 18 0.00 92 0.00 18 0.00 Offset 9 71 0.07 18 0.08 92 0.04 18 0.10 Offset 8 71 0.21 18 0.06 92 0.04 18 0.07 Offset 7 71 -0.50 18 0.01 92 -0.02 18 0.08 Offset 6 71 -0.23 18 -0.10 92 0.01 18 -0.47 Offset 5 71 -0.14 18 -0.32 92 0.19 18 -0.48 Offset 4 71 -0.46 18 -1.82 92 0.03 18 -0.58 Offset 3 71 -0.75 18 -0.79 92 0.04 18 -0.23 Offset 2 71 -0.39 18 -0.04 92 0.03 18 -0.20 Offset 1 71 -0.12 18 -0.18 92 0.35 18 -0.54 Onset 1 49 0.90 16 1.58 73 1.55 17 1.60 Onset 2 49 0.88 16 1.51 73 1.50 17 1.72 Onset 3 49 0.79 16 1.94 73 1.19 17 1.31 Onset 4 49 0.75 16 1.27 73 1.09 17 1.23 Onset 5 49 0.58 16 0.78 73 0.86 17 1.13 Onset 6 49 0.56 16 0.53 73 0.60 17 0.89 Onset 7 49 0.37 16 0.33 73 0.41 17 0.42 Onset 8 49 0.27 16 0.20 73 0.11 17 0.23 Onset 9 49 0.10 16 0.00 73 0.11 17 0.11 Onset 10 48 0.00 16 0.00 74 0.00 17 0.00 REFERENCES Aronson, A. E. (1990). Clinical voice disorders: An interdisciplinary approach (3rd ed.). New York: Thieme. Baken, R. J. (1987). Clinical measurement of speech and voice. Austin, Texas: Pro-Ed. Bielamowicz, S., Kreiman, J., Gerratt, B. R., Dauer, M. S., & Berke, G. S. (1996). Comparison of voice analysis systems for perturbation measurement. Journal of Speech, Language, and Hearing Research, 39(1), 126-134. Boersma, P. & Weenink, D. (2014). Praat: Doing phonetics by computer [Computer program]. Version 5.3.68, retrieved 20 March 2014 from http://www.praat.org/ Boone, D. R., & McFarlane, S. C. (2000). The voice and voice therapy (6th ed.). Boston: Allyn and Bacon. Colton, R. H., & Casper, J. K. (2006). Understanding voice problems: A physiological perspective for diagnosis and treatment (3rd ed.). Baltimore: Lippincott Williams & Wilkins. Crystal, D. (1969). Prosodic systems and intonation in English. Cambridge, U.K.: Cambridge University Press. Dromey, C., Nissen, S. L., Roy, N., & Merrill, R. M. (2008). Articulatory changes following treatment of muscle tension dysphonia: preliminary acoustic evidence. Journal of Speech, Language, and Hearing Research, 51(1), 196-208. Eadie, T. L., & Stepp, C. E. (2013). Acoustic correlate of vocal effort in spasmodic dysphonia. Annals of Otology, Rhinology & Laryngology, 122(3), 169-76. Fairbanks, G. (1960). Voice and articulation drillbook (2nd ed.). New York: Harper & Row. Goberman, A. M., & Blomgren, M. (2008). Fundamental frequency change during offset and onset of voicing in individuals with Parkinson disease. Journal of Voice, 22(2), 178-191. Hillman, R. E., Holmberg, E. B., Perkell, J. S., Walsh, M., & Vaughan, C. (1989). Objective assessment of vocal hyperfunction: An experimental framework and initial results. Journal of Speech and Hearing Research, 32, 373-392. 52 House, A., & Fairbanks, G. (1953). The influence of consonant environment upon the secondary acoustical characteristics of vowels. Journal of the Acoustical Society of America, 35, 84-92. Lien, Y. A. S., Gattuccio, C. I., & Stepp, C. E. (2014). Effects of phonetic context on relative fundamental frequency. Journal of Speech, Language, and Hearing Research, 57(4), 1259-1267. Lien, Y. A. S., & Stepp, C. E. (2013, July). Automated estimation of relative fundamental frequency. In Engineering in Medicine and Biology Society (EMBC), 2013 35th Annual International Conference of the IEEE, 2136-2139. IEEE. Lien, Y. A. S., Michener, C. M., Eadie, T. L., & Stepp, C. E. (2015). Individual monitoring of vocal effort with relative fundamental frequency: Relationships with aerodynamics and listener perception. Journal of Speech, Language, and Hearing Research, 58(3), 566-575. Lofqvist, A., Koenig, L. L., and McGowan, R. S. (1995). Vocal tract aerodynamics in /aCa/ utterances: Measurements. Speech Communication, 16, 49-66. Morrison, M. D., & Rammage, L. A. (1994). The management of voice disorders. San Diego, CA: Singular Publishing Group. Robb, M. P., & Smith, A. B. (2002). Fundamental frequency onset and offset behavior: A comparative study of children and adults. Journal of Speech, Language and Hearing Research, 45(3), 446-456. Roy, N. (2003). Functional dysphonia. Current Opinion in Otolaryngology & Head and Neck Surgery, 11, 144-148. Roy, N. (2008). Assessment and treatment of musculoskeletal tension in hyperfunctional voice disorders. International Journal of Speech-Language Pathology, 10, 195- 209. Roy, N., & Bless, D. M. (1998). Manual circumlaryngeal techniques in the assessment and treatment of voice disorders. Current Opinion in Otolaryngology & Head and Neck Surgery, 6(3), 151-155. Roy, N., Nissen, S. L., Dromey, C., & Sapir, S. (2009). Articulatory changes in muscle tension dysphonia: Evidence of vowel space expansion following manual circumlaryngeal therapy. Journal of Communication Disorders, 42(2), 124-135. Stemple, J. C. (Ed.). (2000). Voice therapy: Clinical studies (2nd ed.). San Diego, CA: Singular. 53 Stepp, C. E. (2013). Relative fundamental frequency during vocal onset and offset in older speakers with and without Parkinson's disease. The Journal of the Acoustical Society of America, 133(3), 1637-1643. Stepp, C. E., Hillman, R. E., & Heaton, J. T. (2010). The impact of vocal hyperfunction on relative fundamental frequency during voicing offset and onset. Journal of Speech, Language and Hearing Research, 53(50), 1220-1226. Stepp, C. E., Merchant, G. R., Heaton, J. T., & Hillman, R. E. (2011). Effects of voice therapy on relative fundamental frequency during voicing offset and onset in patients with vocal hyperfunction. Journal of Speech, Language and Hearing Research, 54(5), 1260-1266. Stepp, C. E., Sawin, D. E., & Eadie, T. L. (2012). The relationship between perception of vocal effort and relative fundamental frequency during voicing offset and onset. Journal of Speech, Language and Hearing Research, 55(6), 1887-1896. Stevens, K. N. (1977). Physics of laryngeal behavior and larynx modes. Phonetica, 34(4), 264-279. Titze, I. R. (1989). Physiologic and acoustic differences between male and female voices. The Journal of the Acoustical Society of America, 85, 1699-1707. Titze I. (1995). Summary statement. Workshop on Acoustic Voice Analysis. Denver. Titze, I. R., Horii, Y., & Scherer, R. C. (1987). Some technical considerations in voice perturbation measurements. Journal of Speech, Language, and Hearing Research, 30(2), 252-260. Van Den Berg, J. (1958). Myoelastic-aerodynamic theory of voice production. Journal of Speech and Hearing Research, 1, 227-244. Watson, B. C. (1998). Fundamental frequency during phonetically governed devoicing in normal young and aged speakers. The Journal of the Acoustical Society of America, 103, 3642-364.