Effects of Deep-Brain Stimulation on Speech: Perceptual and Acoustic Data

Abstract

Purpose:

This study examined speech changes induced by deep-brain stimulation (DBS) in speakers with Parkinson's disease (PD) using a set of auditory-perceptual and acoustic measures.

Method:

Speech recordings from nine speakers with PD and DBS were compared between DBS-On and DBS-Off conditions using auditory-perceptual and acoustic analyses. Auditory-perceptual ratings included voice quality, articulation precision, prosody, speech intelligibility, and listening effort obtained from 44 listeners. Acoustic measures were made for voicing proportion, second formant frequency slope, vowel dispersion, articulation rate, and range of fundamental frequency and intensity.

Results:

No significant changes were found between DBS-On and DBS-Off for the five perceptual ratings. Four of six acoustic measures revealed significant differences between the two conditions. While articulation rate and acoustic vowel dispersion increased, voicing proportion and intensity range decreased from the DBS-Off to DBS-On condition. However, a visual examination of the data indicated that the statistical significance was mostly driven by a small number of participants, while the majority did not show a consistent pattern of such changes.

Conclusions:

Our data, in general, indicate no-to-minimal changes in speech production ensued from DBS stimulation. The findings are discussed with a focus on large interspeaker variability in PD in terms of their speech characteristics and the potential effects of DBS on speech.

Deep-brain stimulation (DBS) is a neurosurgical procedure in which electrodes are implanted into targeted areas in the basal ganglia, most commonly subthalamic nucleus (STN) or globus pallidus internus (GPi; Aldridge et al., 2016; Hartmann et al., 2019). Currently, DBS is an increasingly common and effective surgical treatment option for patients with a wide range of indications, including dystonia, essential tremor, epilepsy, depression, pain, and Alzheimer's disease (Lozano et al., 2019). However, the most common indication for DBS is Parkinson's disease (PD), which is frequently introduced when medication is no longer effective on its own and is typically prescribed as a chronic intervention to obtain continuous stimulation (Herrington et al., 2016; Lozano et al., 2019). Numerous behavioral and neuroimaging studies have provided strong evidence of its clinical benefits in managing negative symptoms of PD, including tremors and bradykinesia, mood and depression, and quality of life (Accolla & Pollo, 2019; Krack et al., 1998; Limousin et al., 1998). Furthermore, medication requirements are reported to diminish substantially with DBS (Kleiner-Fisman et al., 2003). Although direct effects of the stimulation depend on multiple factors such as the stimulation target, dopaminergic treatment, frequency/voltage of stimulation, and stages of PD, DBS has been favored within both research and clinical settings as one of the most important advances in the clinical neurosciences in the past two decades (Lozano et al., 2019). DBS is described as bringing the second honeymoon period (dopaminergic treatment being the first) for patients with PD (Niketeghad et al., 2014).

As such, the scope and number of patients receiving DBS have significantly increased in recent years. Data indicate that every year, up to 4,000 patients with PD and essential tremor receive DBS at 200–250 centers in the United States alone, and the numbers increase yearly (Lozano et al., 2019; Pilitsis et al., 2012). Despite its increasing use in the PD population, DBS is ineffective or even causes adverse reactions for certain symptoms. For example, stimulation of the most common targets (STN, GPi) has been reported to do little for dysarthria in PD or to exacerbate speech and affective/cognitive symptoms (Baudouin et al., 2023; Merola et al., 2017). Rodriguez-Oroz et al. argued that chronic DBS created a new phenotype of PD, characterized by improvement in bradykinesia, tremor, rigidity, on–off fluctuations, and dyskinesia but deterioration in speech, cognition, and gait (Rodriguez-Oroz et al., 2012). A full understanding of the effects of DBS on a variety of motor and nonmotor symptoms remains unavailable.

Speech Outcomes Following DBS in Individuals With PD

With the rapidly growing body of research, speech outcomes following DBS reported in the extant literature on individuals with PD are not easy to summarize, primarily because of the heterogeneity of the methods (i.e., clinical populations, stimulation target/frequency, outcome measures, language studied). For example, target populations have been broadened to include individuals with early motor fluctuations, hence a less advantaged disease stage (Hacker et al., 2020). The most common target is STN (compared to GPi) and bilateral (compared to unilateral) in the clinical standard of PD care (Hartmann et al., 2019). English has been predominantly studied, but research using other languages has also been active (e.g., Czech: Brabenec et al., 2017; Rusz et al., 2023; French: Moreau et al., 2011; Pinto et al., 2014; Japanese: Tsuboi et al., 2015; Quebec French: Martel-Sauvageau & Tjaden, 2017; Martel-Sauvageau et al., 2015, to name a few). All of these factors may have affected the DBS-induced speech outcomes. With this caution, relevant studies are reviewed in the following sections. Unless otherwise marked, the language studied is English. Furthermore, although adverse effects on speech have been reported in both research designs of DBS studies, pre- versus post-DBS and postoperative DBS-On versus -Off (Aldridge et al., 2016), our literature review on speech tasks/stimuli selected in DBS studies is focused on the DBS-On and -Off literature. This is because the current retrospective study compares DBS-On and -Off conditions, as detailed later.

Generally speaking, speech is often resistant to DBS-induced favorable change. According to Wertheimer et al., speech deterioration is a recurrent side effect of STN-DBS, with a prevalence ranging from 4% to 17% of patients across studies (Wertheimer et al., 2014). Tripoliti et al. (2008) reported a significant deterioration in speech intelligibility when patients with PD were stimulated inside or outside the STN, especially at high voltage (4 V), despite improvement in other motor functions. Furthermore, the prevalence of dysarthria under STN-DBS varied from 1% after 6 months of STN-DBS and 70% after 3 years, with an average of 9.3% (Kleiner-Fisman et al., 2006; Obeso et al., 2001; Piboolnurak et al., 2007). An adverse effect on speech has also been reported in patients who underwent DBS for other movement disorders, such as essential tremor (Sandstrom et al., 2020).

Notwithstanding significant interspeaker variability concerning the impact of DBS on voice and speech production in PD (Dromey & Bjarnason, 2011; Hammer et al., 2010; Martel-Sauvageau & Tjaden, 2017; Sidtis et al., 2020; Skodda, 2012; Skodda et al., 2011), a careful review of previous work brings attention to the type and nature of the speech measures employed in these studies. A careful, sophisticated examination of DBS effects on different tasks (ideally those holding better ecological validity by reflecting real-life communication situations) would expand our understanding of DBS effects on speech. In a number of studies including both research designs, pre- versus post-DBS and postoperative DBS-On versus -Off, speech has been assessed and reported as part of the Unified Parkinson's Disease Rating Scale III (UPDRS III) scores (Deep-Brain Stimulation for Parkinson's Disease Study Group, 2001; Kleiner-Fisman et al., 2006; Krack et al., 1998; Limousin et al., 1998; Piboolnurak et al., 2007). The speech item of the UPDRS III rates overall speech function on a 5-point scale (0 = normal, 4 = unintelligible). Although quick and convenient to register, the use of this rank-order rating is often criticized for being inadequate, and further measurements, acoustic or physiological, should complement the ordinal, subjective scale (Gentil et al., 1999). Patients' self-perception has also been included using the Voice Handicap Index (VHI) as well as ratings such as the degree of “slurred speech” or “low volume” (Wertheimer et al., 2014). Some studies used speech intelligibility scores as an index of overall speech function changes across stimulation conditions (e.g., on vs. off, high vs. low voltage, inside vs. above the STN; Sidtis et al., 2020; Tripoliti et al., 2008; Van Lancker Sidtis et al., 2012).

Increasing research has focused on specific aspects of speech production, such as phonation or articulation, to identify the source of the aforementioned overall or global speech deficits. As such, a wide array of acoustic and physiological measures has been examined, including vocal intensity (Tripoliti et al., 2008), the force of the lip and tongue (as measured as ramp-and-hold force contractions; Gentil et al., 2000, 2001; Pinto et al., 2003), displacement and velocity of the lip and tongue (Mücke et al., 2018), long-term average spectrum (LTAS; Tripoliti et al., 2011), and parameters included in automated analysis programs such as the Multidimensional Voice Program and the Advanced Motor Speech Profile (D'Alatri et al., 2008). The findings, in general, have revealed an interesting discrepancy between overall speech function and specific aspects of speech production. That is, despite a decrease in global perceptual ratings post-DBS, more fine-grained, instrumental measurements of speech production showed an improvement in specific measures (i.e., increased vocal intensity and LTAS means, strengthened force of the tongue; Gentil et al., 2000, 2001; Pinto et al., 2003; Tripoliti et al., 2008, 2011). These findings may imply that a set of speech-related measures that bear on the direct relevance of speech function needs to be examined.

A need for careful selection of speech measures to examine the effects of DBS was deliberated by Weismer et al. (2012). The authors first pointed out that many published evaluations of DBS have used relatively crude measures or measures whose relevance to functional speech communication can be questioned. For example, the link between selected measures (e.g., electromyographic recordings from speech muscles, air pressure in the oral cavity) and a metric of functional speech (e.g., speech intelligibility) is largely weak. After reviewing several aerodynamic, acoustic, kinematic, oromotor/nonverbal measures, the authors concluded that an acoustic measure, the slope of second formant (F2), is best suited for the evaluation of DBS due to its ease of application, demonstrated sensitivity to dysarthria and relevance to functional speech communication. This suggestion is consistent with the findings of the studies reporting task-specific effects of DBS on speech. For example, Sidtis et al. (2012) found the adverse effects of DBS on speech intelligibility in conversational speech. However, when the speakers were asked to repeat the same utterances selected from the conversation later on the day of the experiment, intelligibility was not affected by DBS. Based on the findings, the authors argued a facilitative role of external modeling in speech motor control, which was given when the speakers were asked to repeat the selected utterances but not in the original conversational speech (Sidtis et al., 2012). In fact, individuals with PD are well documented to rely on external cues as they have difficulty with skilled movements with only internal cues for gait (Morris et al., 1996), handwriting (Oliveira et al., 1997), and speech production (Weir-Mayta et al., 2017). Taken together, when assessing the impact of DBS on speech outcomes, it is essential to consider the effects of various factors, including instructions and speech tasks.

Several studies have focused on perceptual and acoustic measures, considering their impact on speech intelligibility in dysarthria (Dromey & Bjarnason, 2011; Martel-Sauvageau & Tjaden, 2017; Martel-Sauvageau et al., 2015; Sidtis et al., 2020). A recent study by Sidtis et al. examined changes in speech between on and off bilateral STN-DBS using a series of perceptual evaluations of speech production (Sidtis et al., 2020). The perceptual evaluations included two components: (a) intelligibility as measured by transcription accuracy and listener ratings of the difficulty of each transcription using a scale of 1 (easy) to 5 (difficult) and (b) subject ratings of specific speech characteristics (voice, articulation, fluency, and speech rate). All ratings were obtained using excerpts from a 5-min monologue about a topic of their choice. Similar to other investigations, speech intelligibility was reduced for DBS-On for both high- and low-frequency stimulation conditions (185 and 60 Hz, respectively), compared to DBS-Off. Among the four investigated aspects of speech production, the authors highlighted that STN-DBS exacerbates existing voice problems and may introduce new articulatory problems.

From six patients with PD and bilateral STN DBS, Dromey and Bjarnason (2011) investigated speech changes with DBS turned on and off, using several acoustic measures of articulation (acoustic vowel space, first [F1] and F2 frequency slopes, and a spirantization index) and phonation (perturbation and LTAS) and perceptual (speech severity) ratings. The results, albeit descriptive, because of the small sample size and interpatient heterogeneity, revealed a decrease in speech intelligibility along with a mix of positive and negative speech changes across patients and measures. One finding to note was a reduced vowel space area (VSA) when the stimulation was on for four of the six patients. The authors speculated that the size of vowel space may reflect the restricted mobility of articulators with DBS-On as compared to DBS-Off.

Lastly, in a study investigating DBS effects on vocalic transitions in Quebec French speakers with PD, Martel-Sauvageau and colleagues obtained F2 trajectories for glides and locus equations in consonant–vowel sequences from eight participants who had undergone bilateral DBS of the STN surgery (Martel-Sauvageau & Tjaden, 2017). Both transition metrics, as well as intelligibility scores (passage reading rated by three speech-language pathologists using a visual analog scale), did not show differences between DBS-On and -Off status.

In summary, there is currently general agreement on DBS's overall minimal or negative effects on speech. However, the specific aspects of speech production that may account for the negative changes are still in the early stages of exploration. Furthermore, there is uncertainty about which specific measures are good indices of the adverse DBS effects due to the limited kinds of speech tasks and measures used in this line of work. It seems reasonable to advocate using speech tasks and measures that have reliable relevance to functional speech, especially considering frequent speech deficits in individuals with PD.

The Current Study: A Multilevel Analysis of DBS-Induced Speech Changes

As part of a larger project investigating speech changes in persons with PD who underwent DBS using three measurement domains (auditory-perceptual, acoustic, and kinematic measures), this study reports data on DBS-induced changes between DBS-On and -Off conditions examined by auditory-perceptual ratings and acoustic analyses. The kinematic results on lip and jaw movements are found in the companion paper (Nip et al., 2023). To focus on the effects of DBS on speech, we employed word and sentence stimuli (vs. speechlike tasks such as vowel prolongation or syllable repetition) and measures that have been repeatedly reported as sensitive to speech intelligibility and/or speech deficits related to PD (see the Method section for detail). Although a growing number of studies have included real speech tasks and stimuli, data are still lacking in investigating speech deficits at the fine level. We utilized perceptual and acoustic measures that capture various aspects of speech production (i.e., phonation, articulation, prosody) to describe DBS-induced speech changes. Considering the frequent disturbance in phonation, articulation, and prosody, perceptual ratings of voice quality, articulation precision, and prosody were examined. As an overall speech function index, speech intelligibility was also included. Lastly, the degree of listening effort was rated in consideration of increasing studies supporting its use as an alternative measure of the degree of difficulty in deciphering dysarthric speech (Fletcher et al., 2022; Nagle & Eadie, 2018).

In summary, we posed the following research question, “How does DBS impact the perceptual and acoustic dimensions of speech in individuals with PD?” The primary difference from the existing literature is the use of speech tasks and measures examining real speech and various aspects of speech production.

Method

Participants and Tasks

All participants gave informed consent, and the study protocol was approved by the Florida State University (FSU) and San Diego State University institutional review boards. As a part of a larger project, nine speakers with idiopathic PD and with DBS (PD-DBS: M_age = 72.10, SD_age = 7.32 in years) were included in the study. Participants were randomized as starting with their DBS-On or -Off with a washout period of 30 min between states (Nip et al., 2023; Perera et al., 2015). Acoustic and kinematic signals were recorded after the washout period. Levodopa equivalent daily dose was calculated by referencing the medication list provided by the participants (Tomlinson et al., 2010). Table 1 summarizes participant information.

Table 1.

Demographic information on participants.

Participant	Age (years)	Sex	Time since diagnosis (years)	Time since DBS (years)	LEDD	DBS target	Dysarthria severity
PD-DBS1	60.4	M	17.7	3.75	260	STN-B	Severe
PD-DBS2	69.2	M	20.8	6.25	1,540	STN-B	Moderate
PD-DBS3	69.3	F	11.8	7.33	2,389	STN-B	Mild
PD-DBS4	75.1	F	8.3	1.92	1,666	GPi-L	Mild
PD-DBS5	76.9	M	7.1	0.25	1,150	GPi-B	Mild
PD-DBS6	82.5	F	9.9	1.92	625	GPi-B	Mild
PD-DBS7	69.7	M	14.1	8.75	849	STN-B	Mild
PD-DBS8	64.9	F	11.1	3.00	583	STN-B	Moderate
PD-DBS9	80.9	M	14.2	6.17	200	STN-B	Mild

Note. LEDD = levodopa equivalent daily dose; DBS = deep-brain stimulation; PD = Parkinson's disease; M = male; STN-B = subthalamic nucleus–bilateral; F = female; GPi-L = globus pallidus internus–lateral; GPi-B = globus pallidus internus–bilateral.

Additionally, 44 listeners (M_age = 28.80, SD_age = 5.59 in years) were recruited to provide perceptual ratings. The listener sample included 42 women, one man, and one woman/gender-fluid individual. These listener participants were undergraduate students from the School of Communication Science and Disorders at FSU, with minimal prior exposure to individuals with dysarthria. To be eligible for participation, individuals had to meet the following criteria: (a) native speakers of English and (b) no self-reported history of hearing difficulties. As an incentive, participants received extra credit for their involvement in the study.

Data Collection

Two kinds of speech tasks were selected for the study: word and sentence repetitions. From the database, these tasks were selected as they allow multiple productions of speech stimuli in controlled phonetic contexts. Speakers were seated in front of a large projection screen that showed each speech stimulus and asked to read each stimulus 10 times in a row at their habitual rate and loudness. For recordings, a head-mounted condenser microphone (Shure MC50B) was placed approximately 10 cm from the mouth. The audio signals were recorded by a Marantz digital recorder (PMD660) in .wav format at a sampling rate of 44.1 kHz and with 16-bit quantization.

Table 2 summarizes the stimuli, acoustic measures, and relevant speech components. The words were purposefully selected to include the four corner vowels of English (/i, æ, ɑ, u/) and the diphthong /ɑɪ/. For prosodic characteristics, a relatively long and complex sentence was selected from the database, “The baby birds that saw many butterflies played by the pond” (Kleinow & Smith, 2006).

Table 2.

Summary of the selected acoustic and perceptual measures and speech tasks used for analysis.

Measurement level	Speech dimension	Measure	Stimuli (repetitions)
Acoustic	Articulation	Acoustic vowel dispersion (Hz)	Beet, bat, boot, bot (×10)
		F2 slope (Hz/ms)	Buy Bobby a puppy (×10)
	Prosody	Articulation rate (syl/s)	The baby birds that saw many butterflies played by the pond (×10)
		F0 variation (semitone)
		Intensity variation (dB)
	Phonation	Voiced proportion (%)
Perceptual	Articulation	Articulatory precision (%)	The baby birds that saw many butterflies played by the pond (6th repetition)
	Prosody	Prosody (%)
	Phonation	Voice quality (%)
	Overall	Intelligibility (%)
		Listening effort (%)

Note. F2 = second formant; F0 = fundamental frequency.

The listening experiment was remotely conducted using Qualtrics. The listeners were asked to rate the degree of the five perceptual aspects of speech production: (a) voice quality, (b) articulation precision, (c) prosody, (d) speech intelligibility, and (e) listening effort in response to a sentence. The sentence, “The baby birds that saw many butterflies played by the pond,” was played, and the listener's ratings were collected on a visual analog scale with the words “severely impaired” and “no impairment” marked at the left and right ends, respectively (Sussman & Tjaden, 2012). Considering that speech performance may progressively deteriorate toward the end of the repetition task (Skodda, 2011), we selected the sixth repetition of the production from all speakers.

Modified from Fox and Boliek (2012), ratings were obtained for voice quality, articulation precision, and prosody, considering the frequent speech characteristics of PD (Darley et al., 1969; Fox & Boliek, 2012). Speech intelligibility scores were operationally defined as an index of overall speech severity. Finally, listeners' self-perceived effort in understanding the speech stimuli was included based on the relatively recent literature supporting it as a potential proxy for speech severity, especially when the speakers are mildly dysarthric (Fletcher et al., 2022; Klasner & Yorkston, 2005; Stipancic et al., 2021).

Prior to the perceptual experiment, the listeners completed a brief demographic questionnaire, an audio check, and a brief training on the perceptual task. Due to the remote nature of the perceptual experiment, the delivery of stimuli could not be perfectly controlled. We advised listeners to complete the experiment in a quiet environment with headphones. However, 33 listeners reported using their computer speakers, nine used in-the-ear earphones, and two used over-the-ear headphones. Listeners had the flexibility to adjust the volume to their comfort.

During the training session, the listeners were instructed and practiced performing the ratings in a quiet environment with speech samples presented at a comfortable degree of loudness. As the same sentence was used across the speakers and the conditions, the listeners were explicitly introduced to the sentence and instructed to rate the sentence for the five perceptual aspects.

For each perceptual aspect, listeners evaluated two recordings from the PD DBS speakers, one from the DBS-Off condition and one from the DBS-On condition. Furthermore, the order of the audio files was randomized. For each perceptual dimension, the listeners rated all eight speakers in both DBS conditions, resulting in 16 ratings. Within this set, three of the ratings were randomly chosen to be rated again for investigating intrarater reliability. Consequently, for each perceptual dimension, each listener completed 19 ratings.

Furthermore, the ratings for each perceptual dimension were organized into blocks. For example, listeners provided 19 intelligibility ratings before transitioning to the next perceptual dimension. This approach aimed to reduce potential confusion between dimensions. Before starting each set of ratings, listeners were provided with a description of the specific perceptual dimension and how it differed from those previously rated. The sequence in which perceptual dimensions were rated followed a semirandomized order. The initial set of ratings consistently began with speech intelligibility, followed by listening effort. Subsequent blocks (e.g., articulatory precision, prosody, and voice quality) were randomized to minimize any influence of familiarity on the ratings of intelligibility and listening effort.

Listeners had the option to replay the audio recordings as many times as needed. However, given the substantial number of ratings required in the perceptual experiment, listeners were encouraged to manage their time effectively and avoid spending excessive time on any single recording.

Acoustic Data Analysis

Acoustic measurements were made using Time-Frequency 32 (Milenkovic, 2005) for the following six variables: (a) voicing proportion during utterances, (b) F2 slope, (c) acoustic vowel dispersion, (d) articulation rate, (e) fundamental frequency (F0) range (max–min), and (f) intensity range (max–min; see Table 2). These variables were selected in consideration of their sensitivity to dysarthria, particularly secondary to PD, as well as speech intelligibility in dysarthria (Fischer & Goberman, 2010; Y. Kim et al., 2009, 2011).

For the assessment of articulation, two acoustic measures related to vowels were selected: acoustic vowel dispersion and F2 slope (De Bodt et al., 2002; Y. Kim et al., 2011). Instead of using the traditional VSA measure, this study opted for acoustic vowel dispersion as an alternative. The conventional VSA calculation involves aggregating corner vowel tokens to determine mean F1 and F2 values for each vowel and then using the four aggregated corner vowels to compute the planar area. This approach provides a single VSA value per speaker, which limits the statistical power of the analysis. Consequently, considering the limited number of speakers in this study, acoustic vowel dispersion was employed to overcome the reduced statistical power associated with VSA. Acoustic vowel dispersion was computed by calculating the Euclidean distance between the F1–F2 coordinates of each vowel token and the F1–F2 centroid, which represents the median F1 and F2 values across all vowels (Karlsson & Doorn, 2012; Neumeyer et al., 2010). F1 and F2 were obtained from the temporal midpoint of each vowel (Kuo & Weismer, 2016), and the centroid was calculated separately for the DBS-On versus DBS-Off conditions. As a result, acoustic vowel dispersion measures were obtained for each token of every vowel in the analysis, thus allowing for the examination of the relative degree of centralization for each vowel token.

For F2 slope, F2 trajectories were manually corrected prior to deriving the slopes following the methods by Weismer et al. (1988). Transition onsets and offsets were identified according to the 20 Hz/20 ms rule (i.e., a frequency change of or greater than 20 Hz over 20 ms is operationally defined as transition), and the transition extent (in Hz) was divided by transition duration (in ms) to compute F2 slopes (Y. Kim et al., 2009).

Phonatory and prosodic measures were also included considering the cardinal perceptual characteristics of dysarthria secondary to PD, such as monopitch and monoloudness. The voicing proportion was calculated by dividing the duration of all voicing in the utterance by the total utterance duration and then multiplying the result by 100. When a pause longer than 150 ms occurred in the middle of the sentence, the pause was eliminated from computation (Tsao & Weismer, 1997). This measure was included in the experiment in consideration of the tendency of PD to continue voicing throughout utterances (hence, a greater voicing proportion in PD), possibly reflecting a lack of control in the laryngeal function (Darley et al., 1969).

Reliability

Inter- and intrameasurer reliability were checked for both listener ratings and acoustic measurements. First, for intrarater listener reliability, three recordings within each dimension were randomly selected to be rated again by the listeners at the end of the perceptual experiment. The absolute differences between the listeners' first and second ratings were calculated, and ratings that had differences above or below 2.5 SDs of the mean were removed from the data. This process identified 16 outlying ratings out of the total 3,479 ratings. Following the removal of outliers, the intrarater listener reliability was determined by calculating the relationship between the listeners' first and second sets of ratings while accounting for listener variation. To do this, we constructed a linear mixed-effects (LME) model to predict the listeners' first ratings using their second ratings, with Listener ID as fixed intercepts. The marginal and conditional R² were calculated using the r2_nakagawa function in the performance package in R (Lüdecke et al., 2021) to evaluate the relationship between the listeners' first and second sets of ratings.

For interrater listener reliability, the ratings for each speaker were z-score transformed, and ratings above or below 2.5 SDs of the mean were removed from the data. This process identified 72 outlying ratings out of the remaining 3,463 ratings. In the end, 3,387 ratings were deemed reliable, which were included in the analysis for the study. Following the outlier detection process, the interrater listener reliability was examined using intraclass correlation coefficients (ICCs), calculated in R using the psych package (Revelle & Revelle, 2015). ICCs were employed to evaluate the consistency between several raters (i.e., the average of k raters) by utilizing a two-way random effects model, as described by Koo and Li (2016). The ICCs were calculated for each perceptual dimension (except for listening effort) in DBS-Off and DBS-On conditions. Listening effort was excluded from this analysis, as it reflects the listeners' personal effort while listening to the speaker.

For acoustic measurements, to establish intrameasurer reliability, approximately 10% of the data were reanalyzed by the same measurer who was the first author of this study. The correlation coefficient across the acoustic variables was .998 (p < .01), indicating good intrameasurer reliability. For the intermeasurer reliability, approximately 20% of the data were reanalyzed by a second measurer. The correlation coefficient between the first and second measurers was .984 (p < .01), indicating good intermeasurer reliability.

Statistical Analysis

All statistical analyses were conducted in R (R Core Team, 2023). We created a series of LME models in R using the lmerTest package (Kuznetsova et al., 2017), including one model for the perceptual ratings and six models for the six acoustic measures.

For the perceptual analysis, the data were transformed from wide to long, such that intelligibility, listening effort, articulatory precision, voice quality, and prosody were collapsed into a single measure (i.e., Rating), with a new variable, Dimension, indicating the level of Rating. This transformation was employed to simultaneously examine the DBS effects on the perceptual measures by exploring interaction effects. The perceptual model was specified with the following structure:

$$\text{Rating}={\mathrm{\beta }}_0+{\mathrm{\beta }}_1\times \text{Dimension}+{\mathrm{\beta }}_2\times \mathrm{DBS}$$ (1)

where Rating represents the perceptual rating across the five perceptual dimensions; Dimension represents each level of the Rating measure (i.e., intelligibility, listening effort, articulatory precision, voice quality, and prosody, with intelligibility as the reference); DBS Status represents the DBS condition (i.e., Off vs. On, with Off as the reference); β₀ is the intercept; β₁, β₂, and β₃ are the coefficients associated with the respective predictor variables and their interactions; u represents the random intercepts of Speaker ID and Listener ID; and ε represents the residual error term, accounting for unexplained variability in the outcome. The emmeans package (Lenth et al., 2019) was used to calculate the estimated marginal means to explore the DBS-Off versus DBS-On contrasts at each Dimension level of the Rating outcome.

For the acoustic analysis, we created six LME models to model each acoustic measure using DBS status as a predictor. Each acoustic model was specified with a similar structure:

$$\text{Acoustic\hspace{0.17em}Outcome}={\mathrm{\beta }}_0+{\mathrm{\beta }}_1\times \mathrm{DBS}$$ (2)

However, the model for acoustic vowel dispersion was specified as

$$\text{Vowel\hspace{0.17em}Dispersion}={\mathrm{\beta }}_0+{\mathrm{\beta }}_1\times \mathrm{DBS}$$ (3)

where Acoustic Outcome represents each of the six acoustic measures (i.e., voicing proportion, vowel dispersion, F2 slope, articulation rate, F0 range, and intensity range), DBS Status represents the DBS condition (i.e., Off vs. On, with Off as the reference), β₀ is the intercept, β₁ is the coefficient associated with the respective predictor variable, u represents the random intercepts of Speaker ID (and Vowel for the vowel dispersion model [i.e., beet, bot, boot, bat]), and ε represents the residual error term, accounting for unexplained variability in the outcome.

In total, seven models were constructed for the analysis: one model for the perceptual ratings and six for the acoustic measures. The assumptions of normality of residuals, heteroscedasticity, and linearity were visually inspected using the check_model function in the performance package (Lüdecke et al., 2021). In addition to the predictors outlined above, Speaker Sex was explored as a predictor for the models. However, this predictor was removed from the models during the analysis, as it was not significant and did not significantly improve the model fit. Additionally, the specified models only include random intercepts. Although we were interested in exploring the random slopes, we opted to keep the slopes fixed in order to avoid overspecifying the models with the current study's small sample size. All findings were evaluated at a significance level of α_Bonferroni = .007 to control the family-wise error rate while conducting multiple model comparisons.

Finally, the “Buy Bobby a puppy” recordings for one speaker (PD-DBS 1) were excluded from the analysis due to poor audio quality. Consequently, the F2 slope values could not be computed for this speaker, and this speaker was not included in the analysis for F2 slope, as shown in Figure 2 and Table 7.

Figure 2.

Individual data of deep-brain stimulation (DBS) effects on acoustic measures. F2 = second formant; syl/s = syllables per second; F0 = fundamental frequency.

Table 7.

Statistical results for the six acoustic measures.

Variable	Voicing proportion (%)	F2 slope (Hz/ms)	Vowel dispersion (Hz)	Articulation rate (syl/s)	F0 variation (semitone)	Intensity variation (dB)
Predictors	Estimates	Estimates	Estimates	Estimates	Estimates	Estimates
(Intercept)	68.22 *	3.65 *	443.76 *	4.64 *	11.45 *	28.83 *
DBS-Off	Reference	Reference	Reference	Reference	Reference	Reference
DBS-On	−6.21 *	−0.21	23.94 *	0.22 *	−0.79	−1.49 *
Random effects
σ2	98.03	0.46	13933.46	0.24	12.38	10.11
τ00	422.77 Speaker	1.00 Speaker	7819.05 Speaker	0.46 Speaker	10.28 Speaker	33.07 Speaker
			78853.95 Vowel
ICC	0.81	0.68	0.86	0.65	0.45	0.77
N	9 Speaker	8 Speaker	9 Speaker	9 Speaker	9 Speaker	9 Speaker
			4 Vowel
Observations	180	151	710	195	179	180
Marginal R2 / conditional R2	0.018 / 0.815	0.007 / 0.686	0.001 / 0.862	0.018 / 0.660	0.007 / 0.457	0.013 / 0.769

Note. F2 = second formant; syl/s = syllables per second; F0 = fundamental frequency; DBS = deep-brain stimulation; ICC = intraclass correlation coefficient.

^*p < .007, also indicated in bold font.

Results

Listener Reliability

To examine the intrarater listener reliability, we constructed an LME model to predict the listeners' first set of ratings using their second set of ratings, while accounting for listener variation. The relationship between the first and second sets of ratings was evaluated by examining the marginal and condition R² values. The marginal R² reflects the proportion of the variance of the first ratings that is explained by the variance of the second ratings alone. The marginal R² was .87. The conditional R² reflects the proportion of the variance of the first ratings that is explained by the variance of the second ratings while accounting for speaker and listener effects. The conditional R² was also .87. Taken together, the marginal and conditional R² values indicate good intrarater listener reliability.

To examine the interrater listener reliability, we calculated the ICC for each perceptual dimension for the ratings made in the DBS-Off and DBS-On conditions. The ICC values are reported in Table 3. The interrater listener reliability is interpreted based on the ICC values, where an ICC closer to 1.0 indicates better interrater reliability. The reliability for most dimensions was consistent with previous studies examining VAS ratings of perceptual dimensions (i.e., .70–.80; Stipancic et al., 2023). However, some dimensions had less reliable ratings, such as articulatory precision (ICC_DBS-Off = .55; ICC_DBS-On = .67) and prosody (ICC_DBS-On = .47).

Table 3.

Interrater listener reliability.

Dimension	PD-DBS
	On			Off
	ICC	95% CI	p	ICC	95% CI	p
Intelligibility	.74	[0.52, 0.87]	< .001	.71	[0.48, 0.86]	< .001
Articulatory precision	.55	[0.19, 0.79]	.003	.67	[0.42, 0.84]	< .001
Voice quality	.77	[0.58, 0.89]	< .001	.74	[0.55, 0.88]	< .001
Prosody	.75	[0.55, 0.88]	< .001	.47	[0.05, 0.74]	.016

Note. PD = Parkinson's disease; DBS = deep-brain stimulation; ICC = intraclass correlation coefficient; CI = confidence interval.

Perceptual Results

The descriptive statistics for the five perceptual measures (Intelligibility, Listening Effort, Articulatory Precision, Voice Quality, and Prosody) are displayed by group, sex, and DBS status in Table 4. Additionally, the main findings for the DBS status contrast across the five perceptual measures are depicted in Table 5. Figure 1 demonstrates the changes for each individual. The pairwise comparisons between DBS-On and DBS-Off across the five perceptual measures revealed no significant DBS effects at the Bonferroni-adjusted alpha level of .007.

Table 4.

Descriptive statistics for the perceptual measures.

Perceptual measures	PD-DBS
	Off		On
	M	SD	M	SD
All speakers
Intelligibility (%)	80.12	27.51	78.60	29.90
Listening effort (%)	74.68	32.04	74.20	31.41
Articulatory precision (%)	67.29	32.82	68.85	32.68
Voice quality (%)	66.17	31.70	60.94	33.27
Prosody (%)	63.50	32.90	60.46	33.83
Female speakers
Intelligibility (%)	91.68	13.80	90.08	15.88
Listening effort (%)	82.68	24.44	82.69	24.67
Articulatory precision (%)	74.68	27.60	74.45	27.56
Voice quality (%)	67.28	29.54	63.37	33.16
Prosody (%)	64.51	30.51	63.30	31.03
Male speakers
Intelligibility (%)	73.90	30.88	72.39	33.70
Listening effort (%)	70.35	34.80	69.29	33.84
Articulatory precision (%)	63.15	34.82	65.52	35.05
Voice quality (%)	65.54	32.96	59.54	33.39
Prosody (%)	62.97	34.20	58.85	35.34

Note. PD = Parkinson's disease; DBS = deep-brain stimulation.

Table 5.

Pairwise comparisons of the five perceptual measures between DBS-Off and DBS-On conditions.

Dimension	DBS-Off		DBS-On		Contrast	Estimate	CI	p
	M	SD	M	SD
Intelligibility (%)	80.12	27.51	78.60	29.90	Off–On	1.43	[−2.06, 4.91]	.421
Listening effort (%)	74.68	32.04	74.20	31.41	Off–On	1.25	[−2.23, 4.73]	.482
Articulatory precision (%)	67.29	32.82	68.85	32.68	Off–On	0.59	[−2.87, 4.06]	.737
Voice quality (%)	66.17	31.70	60.94	33.27	Off–On	4.88	[1.32, 8.43]	.007
Prosody (%)	63.50	32.90	60.46	33.83	Off–On	4.64	[1.12, 8.15]	.010

Note. DBS = deep-brain stimulation; CI = confidence interval.

Figure 1.

Individual data of deep-brain stimulation (DBS) effects on auditory-perceptual measures.

Acoustic Results

The descriptive statistics for the six acoustic measures are displayed by group, sex, and DBS status in Table 6. Table 7 summarizes the statistical results between the DBS-On and DBS-Off conditions.

Table 6.

Descriptive statistics (mean and standard deviation) for the six acoustic measures.

Variables	PD-DBS
	Off		On
	M	SD	M	SD
All speakers
Voicing proportion (%)	67.55	20.51	67.06	20.33
Vowel dispersion (Hz)	434.05	292.16	450.69	286.73
F2 slope (Hz/ms)	3.74	1.21	3.50	1.26
Articulation rate (syl/s)	4.73	0.79	4.90	0.90
F0 range (semitone)	10.13	3.66	10.72	3.61
Intensity range (dB)	28.14	6.06	28.43	6.15
Female speakers
Voicing proportion (%)	59.41	19.82	55.19	19.89
Vowel dispersion (Hz)	525.95	369.38	559.27	318.99
F2 slope (Hz/ms)	3.95	1.35	3.57	1.40
Articulation rate (syl/s)	3.99	0.24	4.16	0.42
F0 range (semitone)	11.89	4.13	12.54	4.89
Intensity range (dB)	27.06	6.15	28.00	6.52
Male speakers
Voicing proportion (%)	72.42	19.51	74.18	17.15
Vowel dispersion (Hz)	379.37	217.81	364.87	224.74
F2 slope (Hz/ms)	3.53	1.03	3.46	1.17
Articulation rate (syl/s)	5.15	0.67	5.42	0.78
F0 range (semitone)	9.05	2.88	9.63	1.91
Intensity range (dB)	28.79	5.98	28.69	5.96

Note. PD = Parkinson's disease; DBS = deep-brain stimulation; F2 = second formant; F0 = fundamental frequency.

Among the six acoustic measures, the results of the linear mixed model revealed significant differences for four measures between the DBS-On and DBS-Off conditions. The PD-DBS speakers showed a decreased voicing proportion (p < .001), increased acoustic vowel dispersion (p = .007), faster articulation rate (p = .002), and reduced intensity range (p = .002) in the DBS-On condition compared to the DBS-Off condition. No statistical differences were found for F2 slope and F0 range between the DBS conditions.

Despite the statistical results indicating a significant impact of DBS on the PD-DBS speakers as a group for the four acoustic measures, further visual inspection of individual data indicates large interspeaker variability. The main findings for the DBS status contrast across the acoustic measures are individually depicted in Figure 2.

For example, although voicing proportion during sentence production showed a significant decrease from DBS-On to DBS-Off by 0.49%, it appears that the statistical significance is driven primarily by the speaker PD-DBS 3. Similarly, significant increases in articulation rate by 0.17 syllables per second (syl/s) and decreases in intensity range by 0.29 dB appear to be largely due to PD-DBS 1 and 3, respectively.

Additionally, Figure 3 visualizes acoustic vowel dispersion for each corner vowel. Although acoustic vowel dispersion significantly increased in the DBS-On status as a group, the effect of DBS appears none to negligible for most participants. Its statistical significance appears to be driven by a few participants. The substantial change, either increase or decrease in vowel dispersion for each vowel, appears to lack a consistent pattern within and across speakers.

Figure 3.

Acoustic vowel dispersion for each target word. DBS = deep-brain stimulation.

Discussion

In general, the findings of the study showed none-to-minimal impacts of DBS on speech production in individuals with PD. Particularly, none of the five perceptual ratings showed differences between the DBS-On and -Off conditions. Some acoustic measures significantly changed between the two DBS conditions. Based on the findings, instead of insisting on a general, uniform change induced by DBS in individuals with PD, we argue that the effects of DBS on speech, as a group, are negligible. However, a small subset of individuals with PD may show changes in acoustic signals, either positive or negative, which may not be perceptually captured. Furthermore, such changes may vary substantially across speakers and speech measurements, which are detailed below.

Effects of DBS on Speech Production: Interspeaker Variability

Consistent with previous studies, our results indicate that speech remains largely unchanged between the DBS-On and -Off conditions. Despite no differences in each of the five perceptual ratings, four of six acoustic measures indicated significant changes in the DBS-On condition. Specifically, compared to the DBS-Off condition, speakers decreased voicing proportion during sentence reading while increasing acoustic vowel dispersion, articulation rate, and intensity range. The findings indicate that acoustic measures, when properly selected, may reveal small changes in speech induced by DBS, which are not captured by perceptual ratings, at least using the method of our study.

Caution is required, however, when interpreting the reported significant changes in acoustic signals. As speculated in the results, the statistical significance appears to be driven by a small number of participants. As such, the magnitude of the changes on average is almost negligible: 0.49% (67.55%–67.06%), 16.64 Hz (434.05–450.69 Hz), 0.17 syl/s (4.73–4.90 syl/s), and 0.29 dB (28.14–28.43 dB) from DBS-Off to DBS-On for voicing proportion, vowel dispersion, articulation rate, and intensity range, respectively.

Large interspeaker variability with respect to DBS stimulation effects on speech has been noted likely due to several factors such as target location and electrical parameters (Aldridge et al., 2016; Sidtis et al., 2020). One hypothesis suggests that high-voltage parameters might lead to stimulation outside the surgery target, which could benefit some distal motor symptoms. However, it may result in adverse effects on speech such that DBS patients develop speech problems similar to spastic type of dysarthria. Relatedly, Tsuboi et al. (2015) reported the results of factor analysis and subsequent cluster analysis of speech and voice characteristics that were obtained from 76 Japanese speakers with PD who underwent bilateral STN-DBS (Tsuboi et al., 2015). The results classified the participants into five clusters: relatively good speech and voice function type, stuttering type, breathy voice type, strained voice type, and spastic dysarthria type. The last two voice-speech characteristics (i.e., strained voice and spastic dysarthria) that indicate involvement of the bilateral upper motor neuron system improved after stopping stimulation. Based on the findings, the authors speculated that this speech deterioration in the DBS-On status may be related to current diffusion to the corticobulbar fibers, which also contributes the heterogeneity of DBS effects on speech.

Although further analyses of subsets of the participants were not followed in the study due to the small sample size, our data indicate that speech characteristics of PD vary substantially, which may interact with the DBS-induced changes. Continuous voicing during connected speech is an example. Unlike other acoustic measures included in the study, voicing proportion has been rarely investigated for individuals with PD. However, continuous voicing in PD, especially in severe cases of dysarthria, has been noted in the context of describing perceptual speech characteristics of PD and providing participant exclusion criteria (Duffy, 2012; Fischer & Goberman, 2010). Furthermore, Y. Kim et al. (2011) found shorter voiceless interval durations in individuals with PD but not in other diseases included (traumatic brain injury, multiple system atrophy, and stroke). Based on the findings, the authors suggested shorter voiceless intervals (reflecting extended voicing in voiceless segments and/or relatively normal or fast speaking rate) as a unique feature of hypokinetic dysarthria secondary to PD (Y. Kim et al., 2011).

Our initial hypothesis was that voicing proportion would decrease if laryngeal control improved such that voicing discontinues during voiceless segments. However, another vocal characteristic of PD contradictory to continuous voicing is breathy voice, frequently due to vocal fold bowing (Blumin et al., 2004). In this case, voicing proportion is reduced in PD, and a positive change from DBS would increase voicing proportion, such that vocal closure increases (Hammer et al., 2010). Our PD-DBS group includes participants with the opposite phonatory characteristics, decreased or increased voicing proportion in the DBS-Off condition. Figure 2 shows a bimodal distribution of the voicing proportion data, which is further evident considering that the average value of voicing proportion for healthy controls is approximately 65% from the sentence stimulus (see Table 6). Participants showed little changes between the DBS conditions for both groups at the low and high end of the distribution. However, two participants with increased voicing proportion, PD-DBS 3 and 6, exhibited a substantial decrease in the DBS-On condition (see Figure 2).

Substantial variability among patients in the overall clinical expression of PD is well recognized, and increasing efforts have been made to propose subtypes of PD (Paulus & Jellinger, 1991; Stebbins et al., 2013; Wolters, 2008). Although speech deficits secondary to PD are typically discussed using the label hypokinetic dysarthria, a wide spectrum of speech characteristics in PD has also been noted, possibly reflecting factors such as subtypes of PD and sex. That is, those who are classified as having postural instability and gait disturbance (PIGD) tend to develop more speech deficits than tremor dominant type or indeterminant subtypes (Stebbins et al., 2013). Because of the higher frequency of males presenting with a PIGD type, males tend to exhibit more speech problems than females (R. Kim et al., 2018). It is possible that speech profiles of PD may vary in divergent underlying mechanisms. In fact, a recent study suggested three speech subtypes in PD, prosodic, phonatory-prosodic, and articulatory-prosodic, based on the speech characteristics of 111 participants with de novo PD (Rusz et al., 2021). Speech changes resulting from DBS may be further intricate due to large interspeaker variability regarding responses to medication or behavioral treatment (Huber et al., 2003; Knowles et al., 2018; Rusz et al., 2021). It is beyond the scope and capability of the current study to identify factors of the variability of speech characteristics and its effects on the speech outcomes of DBS. However, the findings support different speech characteristics of individuals with PD, which should be considered part of an individualized, patient-oriented decision when DBS is available as a treatment option.

Consideration of Speech Materials and Measures for DBS Studies

One motivation of the study was the use of speech tasks (e.g., word and sentence production) in examining DBS effects on speech. This was based on the observation that prior work has mostly used speechlike tasks (e.g., vowel prolongation, syllable repetition) to investigate DBS-induced changes in various aspects of speech (e.g., respiratory, phonatory, and articulation). As a retrospective study, we selected speech recordings from an archived database primarily based on acoustic measures of interest. As such, our speech stimuli were, in general, relatively short and simple. For example, one-syllable words (e.g., beet, bat) were used for the acoustic dispersion of the four corner vowels. Furthermore, F2 slope was computed from the word “buy” in the sentence, “Buy Bobby a puppy.” Considering the extant literature supporting the effects of speech stimuli and tasks, it is possible that our speech stimuli and tasks were not optimal in revealing speech disturbances of PD. That is, the length and phonetic complexity of speech stimuli are known to be important factors regarding the degree to which the stimuli are sensitive to speech characteristics of PD (Van Lancker Sidtis et al., 2010, 2012). Similarly, the target word positioned at the beginning of a short sentence was possibly produced with prominence, which consequently strengthens its acoustic and articulatory characteristics (Fougeron & Keating, 1997; Thompson, 2023). The use of extended and complex speech stimuli (e.g., connected speech) may have shown a better sensitivity to PD speech.

Additionally, prior work examining DBS effects on speech has often employed the measures of global speech symptoms and/or the impact of speech on social interaction, frequently using the patients' self-evaluation. This includes the VHI and other self-evaluation questionnaires, such as, “To what extent do you think other people can understand you?” and “Do you socialize less due to speech difficulties?” (Jacobson et al., 1997; Wertheimer et al., 2014). Speech intelligibility has been increasingly included as an overall functional index of stimulation-induced changes (Knowles et al., 2018; Tripoliti et al., 2014; Van Lancker Sidtis et al., 2012). In addition to speech intelligibility, we recommend including additional measures that evaluate subsystems of speech production (e.g., phonation, articulation) for a comprehensive profile of speech production. This recommendation is based on the fact that our perceptual ratings showed a varying degree of speech impairments for both DBS-On and -Off conditions, as demonstrated in Table 3 (voice quality = prosody < precise articulation < listening effort < speech intelligibility from the lowest to the highest ratings).

Furthermore, pertinent to our previous discussion regarding large speaker variability, our participant PD-DBS 2 showed a substantial decrease in two vowel articulation-related measures, F2 slope and acoustic vowel dispersion, despite the lack of changes in other acoustic measures (see Figure 2). Sidtis et al. raised the possibility that DBS (specifically STN-DBS) exacerbated existing voice problems while introducing new articulatory deficits (Sidtis et al., 2020). Taken together, it appears important to consider multiple aspects of speech production in order to develop a comprehensive profile of DBS-induced speech changes.

Limitations and Future Directions

The limitations of the current study include the small number of participants and the limited kinds of speech stimuli. Considering the large interspeaker variability regarding speech characteristics in the DBS-Off condition as well as speech changes induced by DBS, a large-scale study that allows an analysis of subsets of the participants is warranted. The DBS setting (e.g., stimulation target, bilateral/unilateral) was not the same across our participants. Furthermore, the type of speech stimuli and tasks needs to be expanded in future studies to consider variables such as phonetic complexity and position effects. A follow-up analysis is under progress in which alternative measures are investigated, including the deteriorating tendency over repetitions of utterances.

Conclusions

This study examined the effects of DBS on word and sentence production in individuals with PD using auditory-perceptual and acoustic measures. The findings add further data supporting minimal changes in speech induced by DBS as well as large interspeaker variability in acoustic changes. Based on the findings, we highlight the need for consideration of individual speech characteristics. Additionally, the collection of data from more individuals with PD who underwent DBS will assist in determining trends in a comprehensive aspect of speech with large variability demonstrated in this study.

Author Contributions

Yunjung Kim: Writing – original draft (Lead), Conceptualization (Lead), Formal analysis (Lead), Resources (Equal). Austin Thompson: Writing – original draft (Supporting), Formal analysis (Equal). Ignatius S. B. Nip: Conceptualization (Supporting), Writing – review & editing (Supporting), Resources (Equal).

Data Availability Statement

Due to the nature of the study, the speech recordings generated and analyzed for this study are not publicly available to be compliant with the institutional review board requirements. However, data spreadsheets with de-identified participant information may be available from the first author upon reasonable request.

Funding Statement

The study was in part supported by three research funds: National Institutes of Health (NIH) R01 DC 020468 and Korea Health Industry Development Institute HI22C0736 (principal investigator [PI]: Yunjung Kim) and NIH F31 DC020121 (PI: Austin Thompson). Part of the data was presented at the Annual Convention of the American Speech-Language-Hearing Association in 2022 (New Orleans, LA).