An Acoustic Study of the Relationships Among Neurologic Disease, Dysarthria Type, and Severity of Dysarthria

Abstract

Purpose:

This study examined acoustic predictors of speech intelligibility in speakers with several types of dysarthria secondary to different diseases and conducted classification analysis solely by acoustic measures according to 3 variables (disease, speech severity, and dysarthria type).

Method:

Speech recordings from 107 speakers with dysarthria due to Parkinson’s disease, stroke, traumatic brain injury, and multiple system atrophy were used for acoustic analysis and for perceptual judgment of speech intelligibility. Acoustic analysis included 8 segmental/suprasegmental features: 2nd formant frequency slope, articulation rate, voiceless interval duration, 1st moment analysis for fricatives, vowel space, F0, intensity range, and Pairwise Variability Index.

Results:

The results showed that (a) acoustic predictors of speech intelligibility differed slightly across diseases and (b) classification accuracy by dysarthria type was typically worse than by disease type or severity.

Conclusions:

These findings were discussed with respect to (a) the relationship between acoustic characteristics and speech intelligibility and (b) dysarthria classification.

One of the most important developments in the understanding of dysarthria was the introduction of a classification system by Darley, Aronson, and Brown (1969a, 1969b). This system (hereafter, the Mayo System) is widely used for research and clinical purposes, but questions persist concerning its reliability and validity for different groups of raters. Zyski and Weisiger (1987), who reported on ratings by graduate students and experienced clinicians, expressed doubts about the suitability of the system for clinical purposes, noting low reliability and low rates of accurate classification. More recently, Fonville et al. (2008) reported on classification accuracy for neurologists and neurology trainees, and Van der Graaf et al. (2009) reported on classification accuracy for neurologists, residents in neurology, and speech therapists. In these studies, the rate of correct classification ranged from about 35% to 40%. Both Fonville et al. and Van der Graaf et al. concluded that classification by perceptual judgment alone is not adequate and that professionals should rely on other sources of information. On the other hand, Guerra and Lovely (2003) reported promising results for classification of dysarthria type using nonlinear self-organizing maps operating on a combination of acoustic and perceptual data. A recent study by Liss et al. (2009) reported that rhythm metrics are able to distinguish dysarthric speech from controls and even between dysarthria subtypes, when the subtypes are chosen to be highly representative of their classical description. Forty years after publication of the Mayo System, it seems fair to say that its classification accuracy, either by perceptual or instrumental measures, has not been adequately established in the research literature.

Severity of Dysarthria

A potentially complicating factor in the classification of dysarthria type is speech severity. The Mayo Clinic studies that provided the empirical basis for classification of dysarthria did not control for severity. Rather, severity was allowed to vary within each of the disease groups studied (Darley, Aronson, & Brown, 1975). There is no standard measure of speech severity in dysarthria, but estimates of speech intelligibility are often used to index the extent to which neurological disease affects the speech mechanism (R. D. Kent et al., 1989). The main barrier to separating effects of severity from types of dysarthria, or other potential classification variables (see the Discussion section), has been the lack of relevant analyses from a sufficiently large number of speakers with different dysarthria types and various levels of severity. The present study draws from a relatively large dysarthria database generated by collaboration between the Mayo Clinic and the University of Wisconsin—Madison to investigate the distribution of selected speech acoustic characteristics within and across dysarthria types and levels of severity, the latter based on estimates of speech intelligibility.

Reviews of dysarthria symptoms at both the articulatory and acoustic levels of analysis (Weismer, 1997; Weismer & Kim, 2010) identify much that is common to speakers with dysarthria, regardless of type. Moreover, the typically large interspeaker variability observed in almost any study of a particular dysarthria type is likely due, in part, to variations in severity of speech involvement. Even in the original Mayo Clinic studies (Darley et al., 1969a, 1969b), there was substantial overlap across dysarthria types in perceptual characteristics (see discussion in Weismer, 1997). Possibly, variation in speech severity within a dysarthria type explains as much variance in physiological, acoustic, and/or perceptual data as variation across dysarthria type. If this is the case, classification of speakers according to speech severity level might be expected to be roughly as accurate as classification by dysarthria type, when properly selected acoustic measures are used as input variables.

The issue of classification accuracy can be expanded to include other potential classification variables. One such variable is disease type. In the original Mayo Clinic studies, the groups were formed based on disease or groups of related diseases such as Parkinson’s disease, cerebellar disease, and so forth (Darley, Aronson, & Brown, 1975). In many cases, there is good correspondence between disease type and dysarthria type, but in other cases, the correspondence is weaker, with multiple dysarthria types being possible for a single disease type because some diseases can affect more than one component of the motor system (Duffy, 2006, pp. 31–33, Table 2–3 and associated text). In the current study, classification according to disease type was also investigated. A reasonable initial hypothesis derived from the Mayo System and its presumed neuropathological basis is that classification according to dysarthria type will be better (i.e., more accurate) than classification based on disease type or severity of speech involvement. This prediction is reasonable because a single disease may be associated with more than one dysarthria type, as noted previously, and severity plays no classification role in the Mayo System.

Table 2.

Results of regression analysis between acoustic variables and scaled speech intelligibility for subjects across all disease groups.

Variable	Gender/Window	r 2
Vowel space	M	.28*
	F	ns
F2 slope	M	.51**
	F	.46**
M1 difference for /s/ – /∫/	1st window	.34**
	2nd window	.33**
	3rd window	.34**
Voiceless interval duration		.30**
Articulation rate		.56**
F0 interquartile range		.38**
Intensity interquartile range		ns
Pairwise Variability Index		ns

^*p < .05.

^**p < .01.

Table 3.

Classification analysis result for each single acoustic variable.

%	F2		VS		M1	VID	AR	dB	F0	PVI
Etiology	All	36.3	All	35.0	38.6	51.0	49.5	33.0	33.0	32.0
	M	47.6	M	17.2
	F	56.4	F	38.5
Type	All	36.3	All	22.3	37.6	17.6	27.2	19.4	31.1	19.4
	M	34.9	M	23.4
	F	53.8	F	20.5
Severity	All	58.8	All	45.6	58.4	46.1	51.5	42.7	41.7	32.0
	M	66.7	M	50.0
	F	56.4	F	46.2

Note. Numbers in the table indicate the number of subjects who were correctly identified by discriminant anlaysis as original categories. For second formant slope (F2) and vowel space (VS), results are reported with separation between male and female speakers. Each acoustic variable is indicated in abbreviated form. M1 = first-moment difference; VID = voiceless interval duration; AR = articulation rate; dB = intensity range; F0 = fundamental frequency range; PVI = Pairwise Variability Index.

The classification issue in dysarthria is important for several reasons. First, a complete understanding of any disorder and the development of a scientific program to investigate its characteristics depend on a sound theoretical basis. The theoretical basis of the distinctions between dysarthria types is an assumed feature of the Mayo System but has rarely been investigated when other classification options (such as severity or disease type) are considered. In other words, there is little empirical basis for the classification scheme specified by Darley et al. (1969a, 1969b), as compared with other potential classification approaches. Challenges to the Mayo System based on perceptual judgments have been described previously in this article. Second, investigations in dysarthria often involve groups of participants classified as having a common dysarthria and who are compared with a control group. Much more rarely, multiple groups of participants with perceptually judged, varying dysarthria types may be compared with each other (and with a control group) for distinguishing characteristics (e.g., Liss et al., 2009). Investigation of alternate classification schemes may inform scientists concerning the relative homogeneity of participant groups defined in different ways.

Acoustic Methods

The current classification analyses were based on acoustic measures thought to reflect salient aspects of speech production in persons with motor speech disorders (Kent & Kim, 2008). The advantage of an acoustic approach to understanding motor speech disorders has been noted (Ansel & Kent, 1992; Weismer, 1984). Previous efforts have pursued two main goals: (a) identification of the signal properties underlying intelligibility deficits in dysarthria and (b) identification of acoustic characteristics of specific dysarthria types. Studies have been conducted to (a) identify acoustic measures that predict speech intelligibility scores and (b) investigate the physical correlates of perceptual features in selected types of dysarthria. Examples of acoustic parameters that predict speech intelligibility in speakers with dysarthria include acoustic vowel space (Liu, Tseng, & Tsao, 2000; McRae, Tjaden, & Schoonings, 2002; Tjaden & Wilding, 2004; Weismer, Jeng, Laures, Kent, & Kent, 2001); second formant frequency (F2) slope (J. F. Kent et al., 1992; R. D. Kent et al., 1989; Kim, Weismer, Kent, & Duffy, 2009; Mulligan et al., 1994; Weismer, Martin, Kent, & Kent, 1992); and voice onset time (VOT; Liu et al., 2000).

In addition, acoustic measurements have been used to investigate characteristics of specific types of dysarthria. For example, ataxic dysarthria has been characterized by slow speaking rate, relatively great variability in VOT, a tendency toward equalized vowel/syllable durations within utterances, and an unusually large fundamental frequency (F0) range across utterances (Ackermann & Hertrich, 1997; Chiu, Chen, & Tseng, 1996; R. D. Kent et al., 2000; Stuntebeck, 2002). Salient acoustic features of hypokinetic dysarthria, as another example, have included either normal or faster-than-normal speaking rates, relatively high mean F0, decreased F2 extents and slopes, and decreased F0 variability (Canter, 1963; Forrest, Weismer, & Turner, 1989; Goberman, Coelho, & Robb, 2005; Solomon & Hixon, 1993; Weismer, 1984, 1991). Other studies have investigated selected acoustic characteristics in spastic (Ozawa, Shiromoto, Ishizaki, & Watamori, 2001; Ozsancak, Auzou, Jan, & Hannequin, 2001; Ziegler & von Cramon, 1986), hyperkinetic (Ackermann, Hertrich, & Hehr, 1995; Hertrich & Ackermann, 1994; Ludlow, Connor, & Bassich, 1987), and flaccid (Morris, 1989) dysarthria, as well as in mixed dysarthrias (Liss et al., 2009; Wang, Kent, Duffy, & Thomas, 2005).

Interestingly, although the Mayo System is widely accepted in clinical practice and for the definition of presumably homogeneous participant groups in research studies (see Duffy, 2005), very few studies other than Liss et al. (2009) have examined the possibility of classification of dysarthria on the basis of acoustic attributes. The unique clusters of perceptual characteristics thought to be the core of the Mayo System (Darley et al., 1975) prompt the hypothesis that properly selected and combined acoustic measures can discriminate among dysarthria types with a reasonable degree of accuracy. As described previously, the same hypothesis of classification from acoustic variables can be entertained for the classification variables of severity and neurological disease type (and, possibly, other variables not discussed here).

Apart from Liss et al. (2009), there is no study of classification of dysarthria, in which a single protocol has been used with a relatively large number of speakers having different dysarthria types, varying severities of speech involvement, and different underlying diseases. In general, studies in which instrumental measures of any type have been used to differentiate the dysarthrias are relatively rare and, when available, are usually based on a single measure (or derivatives of that measure). For example, Nishio and Niimi (2001) measured speaking rate and measures associated with it (articulation rate, pause time) in seven different dysarthria types; these measures were not particularly effective in distinguishing among the different dysarthrias. Similarly, Morris (1989) showed long-lag VOTs to be shorter than normal in talkers with five different types of dysarthria. Formant frequency measures, formant transition rates, and even lip/jaw motions have all been shown to be similar across different types of dysarthria (see Ackermann & Hertrich, 1997; see also reviews in Weismer, 1997; Weismer & Kim, 2010). It is unknown, however, whether combinations of measures reflecting different aspects of articulatory behavior (or the speech acoustic signal) would distinguish among different dysarthria types.

To begin the process of exploring the classification of dysarthria based on multiple acoustic measures, the following questions were asked. First, are some or all of the selected acoustic measures correlated with a measure of speech severity, which in this case was operationalized with a speech intelligibility measure? Second, how well do combinations of acoustic measures classify (a) dysarthria types, (b) disease type, and (c) speech severity as indexed by speech intelligibility measures?

Method

Speakers

One hundred and seven subjects with dysarthria consequent to Parkinson’s disease (males [M] = 29, females [F] = 10), stroke (M = 21, F = 18), multiple system atrophy (M = 11, F = 6), and traumatic brain injury (M = 7, F = 5) were selected for the present study from the University of Wisconsin—Madison Mayo Clinic dysarthria database, which consists of digital speech recordings obtained at the Mayo Clinic in Rochester, Minnesota. Parkinson’s disease, stroke, and traumatic brain injury groups were chosen for this study because they are the most frequent etiologies associated with dysarthria in the United States (Centers for Disease Control and Prevention, 2003; Duffy, 2005; National Institute of Neurological Disorders and Stroke, 2001; National Stroke Association, 2002); the group of speakers with multiple system atrophy was chosen because of the availability of a relatively large number of speakers with this diagnosis and the association of this disease with different types of dysarthria. Across all groups of speakers, participants ranged between 20 and 91 years of age (Mdn = 64.5 years). The dysarthria diagnoses were made by a widely acknowledged expert in the area (Duffy, 2005) who was aware of each patient’s medical diagnosis (if it was known at the time of the classification) but who identified the dysarthria type largely, if not entirely, on perceptual evaluation of a patient’s speech (J. R. Duffy, personal communication, June 2, 2010). The dysarthria types he identified included ataxic, spastic, hypokinetic, flaccid, hyperkinetic, Unilateral Upper Motor Neuron (UUMN), and mixed. Table 1 shows subject information, including the distribution of dysarthria types for these 107 participants; the mixed category pools that the different combinations (e.g., spastic–flaccid, spastic–ataxic) used in the Mayo System. Potential participants with language disorders, apraxia of speech, or aprosodia were excluded from the current study.

Table 1.

Subject information including age, gender, dysarthria type, and scaled speech intelligibility.

Disease	Gender	Age range and median (in parentheses)	Dysarthria type	Speech intelligibility
PD	M = 29 F = 10	38–81 (68)	Hypokinetic (n = 34) Mixed (n = 5)  Hypokinetic–hyperkinetic: 3  Hypokinetic–spastic: 1  Hypokinetic–ataxic: 1	Mild =19 Moderate = 16 Severe = 4
Stroke	M = 21 F= 18	23–91 (71)	UUMN (n =18) Spastic (n = 4) Flaccid (n = 3) Ataxic (n = 2) Mixed (n = 12)  Ataxic–UUMN: 2  Ataxic–spastic: 5  Spastic–flaccid: 1  UUMN–spastic: 1  Ataxic–UUMN–spastic: 2  UUMN–spasitc–hypokinetic: 1	Mild = 16 Moderate = 14 Severe = 9
TBI	M = 7 F = 5	20–51 (33.5)	Flaccid (n = 2) Hyperkinetic (n = 1) Mixed (n = 9)  Ataxic–hypokinetic: 3  Ataxic–spastic: 1  Flaccid–spastic: 1  Spastic–hypokinetic: 1  Ataxic–spastic–hyperkinetic: 1  Ataxic–hyperkinetic–spastic–flaccid: 1	Mild = 6 Moderate = 2 Severe = 4
MSA	M= 11 F = 6	46–71 (59)	Ataxic (n = 7) Hypokinetic (n = 3) Spastic (n = 1) Mixed (n = 6)  Spastic–ataxic: 3  Hypokinetic–spastic: 2  Ataxic–hypokinetic–spastic: 1	Mild = 10 Moderate = 5 Severe = 2

Note. PD = Parkinson’s disease; TBI = traumatic brain injury; MSA = multiple system atrophy; UUMN = Unilateral Upper Motor Neuron; M = male; F = female.

Procedure

The speech samples used in this study included word and sentence recitations. Participants were asked to produce each of six words (hail, shoot, sigh, sip, ship, and wax) 10 times in a row. These words were selected because of their acoustic characteristics and previous demonstrations of their sensitivity to varying severities of dysarthria (R. D. Kent et al., 1989; Weismer, Kent, Hodge, & Martin, 1988). The vocalic nuclei of some of these words require relatively extensive changes in vocal tract shape that were consistent with the sensitivity of formant transitions to speech symptoms in dysarthria (Kim et al., 2009). Other words were chosen for a specific type of analysis appropriate to analysis of consonant production (e.g., sip vs. ship for first-moment (M1) analysis; acoustic measures are explained later in this article). Among the 10 repetitions for each word, the middle eight of the repetition string were taken for analyses to eliminate possible initial and final location effects.

For the sentence recitation task, all participants recited the following five sentences in response to a live-voice demonstration by the examiner: (a) “ Put the high stack of cards on the table,” (b) “Combine all the ingredients in a large bowl,” (c) “ The blue spot is on the key,” (d) “ The potato stew is in the pot,” and (e) “ The boiling tornado clouds moved swiftly.” These sentences were used to obtain speech intelligibility data as well as acoustic data. Three listeners who had a background in speech-language pathology, but who were not expert listeners of speech disorders in general or dysarthria specifically (in the sense defined by Monsen, 1983, for listeners of speakers with profound hearing impairment), made judgments of speech intelligibility. Intelligibility data were obtained using a direct magnitude estimation technique (Gescheider, 1976). Inter- and intralistener variability in selection of number scale ranges was eliminated by following a procedure described by Engen (1971).

The speech samples were collected in a quiet room with a high-quality microphone (SHURE SM 58) and a digital audiotape recorder (DAT; TASCAM DA-P1) at a sampling rate of 44.1 kHz and with 16-bit quantization. After the utterances had been recorded on DAT, they were analyzed using the speech analysis program TF32 (Milenkovic, 2001).

All experimental procedures were approved by the University of Wisconsin—Madison Human Subject Committee.

Acoustic Analysis

Acoustic measurements were made of (a) sentence duration, (b) vowel duration, (c) voiceless interval duration, (d) first and second formant frequencies from four corner vowels, (e) M1 for fricatives (/s/ and /∫/) during three 50-ms-long windows approaching the vocalic nucleus (25-ms overlap between adjacent windows), (f) transition duration and extent for F2 transitions, (g) F0 contour, and (h) root-mean-square (RMS) intensity contour. Among these measures, the voiceless interval duration was used as measured; the remaining measures were used to derive the following variables for further statistical analysis: (a) articulation rate, (b) Pairwise Variability Index (PVI), (c) acoustic vowel space, (d) M1 difference between /s/ and /∫/, (e) F2 slope, (f) F0 range (maximum–minimum) of utterance, and (g) RMS intensity range of utterance. Except for the voiceless interval duration, these acoustic variables were selected based on previous studies that have reported them to be useful either for predicting intelligibility scores or characterizing the production deficit in dysarthric speech. Voiceless interval duration was investigated as an alternative for VOT because of recent concerns about the interpretation of the latter measure (Auzou et al., 2000) and its sensitivity to age-related effects on speech production (Weismer & Fromm, 1983). The PVI served as an estimate of the degree of scanning speech (see Low, Grabe, & Nolan, 2000; Stuntebeck, 2002). Word materials were used to derive M1 for the fricative pair (sip vs. ship) and F2 slopes from vocalic nuclei (hail, wax, sigh, and shoot), whereas sentence duration, vowel duration, voiceless interval duration, acoustic vowel space, overall F0, and RMS intensity range were measured from sentence materials.

Tracking errors observed for formant trajectories and F0 values (usually due to poor voice quality or sudden phonation change) were manually modified using the interactive editor in TF32. Considering possible outliers in F0 and RMS intensity contours, the interquartile range was used to estimate the range of variation for F0 and intensity contours.

Ten subjects were randomly selected in order to examine intrajudge reliability for acoustic measurements. The correlation coefficient computed between the original and remeasured data (after 9 months) across all acoustic variables was .95, which suggested that the measurements were repeatable and were made under a replicable set of criteria.

Results

Speech Intelligibility Scores

The modulus-equalized speech intelligibility score across the 107 speakers with dysarthria ranged from 0.58 to 1.63 (Mdn = 1.40). If listeners were actually doing ratio scaling, this suggests that intelligibility among these speakers varied over a roughly 2.8:1 range (i.e., the most intelligible speaker was scaled as approximately 2.8 times more intelligible than the least intelligible speaker). Speech intelligibility scores for individual disease groups are displayed in Figure 1. The Parkinson’s disease group had the highest mean intelligibility of the four groups, and the stroke group had the lowest; t test results (adjusted per comparison for an overall alpha level of .05) revealed that there were no significant differences in speech intelligibility among the clinical groups.

Figure 1.

Box-and-whisker plot of intelligibility scores, plotted as modulus-equalized values, for speakers with dysarthria in the four disease groups. Mean values of each group are indicated by the dotted line within the box and median by the solid line. The left (lower value) edge of each box is the 25th percentile of the distribution; the right (upper value) edge is the 75th percentile; the whiskers show the lowest and highest nonoutlier observations, and the individual plotted points are outliers. PD = Parkinson’s disease; TBI = traumatic brain injury; MSA = multiple system atrophy.

To estimate intrajudge reliability, each listener scaled 20 randomly selected samples twice. Kendall’s tau-b correlation analyses reported significant interrater correlation (r = .77, p < .01) and intrarater correlation (r = .71, p < .01).

Acoustic Measures Regressed Against Speech Intelligibility

Relationships between speech intelligibility and acoustic measures were determined by regression analysis. Regression analysis was conducted twice for each acoustic variable. One set of analyses examined the general relationships of acoustic variables to speech intelligibility in speakers with dysarthria, pooled across disease groups. The second set examined the same relationships within each of the four disease groups (the regression analyses were restricted to disease types; see the Discussion section). The results of this study suggested that six of eight acoustic measures were significantly correlated with intelligibility when data were pooled across all disease groups (see Table 2). These variables include vowel space, F2 slope, M1 difference between /s/ and /∫/, voiceless interval duration, articulation rate, and F0 range. All measures assumed to reflect segmental articulation (vowel space, F2 slope, M1 difference for fricatives, and voiceless interval duration) were regressed significantly on speech intelligibility scores, as were two of four measures assumed to reflect prosodic features (articulation rate and F0 range). Figures 2 and 3 display examples of significant relationships of F2 slope and articulation rate with scaled speech intelligibility.

Figure 2.

The relationships between second formant frequency (F2) slope of the word wax and speech intelligibility scores: male (M) speakers with dysarthria (left) and female (F) speakers with dysarthria (right). **p < .01.

Figure 3.

Scatter plot of articulation rate and speech intelligibility scores. **p < .01.

It is noteworthy that some acoustic variables appeared to make a significant contribution to speech intelligibility regardless of disease, whereas other variables seemed to be more strongly related within a particular disease group (see Table 2). For example, F2 slope was regressed significantly on speech intelligibility scores for all four disease groups, whereas articulation rate was significant for all clinical groups except PD, and F0 range was significant only for the stroke group. Not surprisingly, t test results indicated that significant acoustic predictors of speech intelligibility common to all four disease groups (e.g., F2 slope) showed no significant differences among diseases, whereas those predictors specific to a certain disease (e.g., articulation rate) showed significant differences among diseases.

Classification Analysis

For classification purposes, participants were grouped by (a) disease, (b) dysarthria type, and (c) speech intelligibility level. The goal of the classification analysis was to answer the second research question: Which grouping, if any, results in the best prediction of group membership by speech acoustic measures? In theory, classification of patients under any of the three groupings could be non-systematic, suggesting that the predictor variables failed to produce a significant discriminant function because the variables were poorly chosen, the grouping variable was poorly conceived, or as a result of some combination of the former and latter. On the other hand, significant classification functions for all groupings would suggest that any one or all of the grouping variables have theoretical and/or clinical relevance.

Subjects were coded according to their disease, dysarthria type, and severity as follows. Disease coding was made following medical records, and dysarthria type was coded based on the perceptual judgment of an acknowledged expert as described in the Method section; this expert classification resulted in seven dysarthria types (ataxic, flaccid, spastic, hypokinetic, hyperkinetic, UUMN, and mixed [the latter as defined previously]). For severity coding, subjects were divided into three groups of operationally defined, nonoverlapping levels of speech intelligibility: mild, moderate, and severe. The severe group was defined as speakers with speech intelligibility scores from 0.58 to 1.17 (Mdn = 1.09, n = 19), the moderate group was from 1.22 to 1.39 (Mdn = 1.34, n = 37), and the mild group was from 1.40 to 1.63 (Mdn = 1.51, n = 51).

Classification analysis was conducted twice using SPSS Version 16.0 for each single acoustic variable and for all eight acoustic variables. The classification results for each single acoustic variable are displayed in Table 3. The values in the table cells indicate the percentage of subjects who were classified identically with their original group. For example, when subjects were classified solely by articulation rate, 49.5% of the subjects were classified identically with their original etiology groups, 27.2% of the subjects with their original dysarthria types, and 51.5% with their original severity groups. Among the eight acoustic variables, seven showed the greatest number of subjects classified identically when severity was the classification variable. When all eight acoustic variables were combined for the discriminant function, however, the results were slightly different (see Table 4). When the stepwise discriminant function was performed by pooling male and female talkers, subjects were significantly classified with their original group by etiology (68.6%), F(6, 190) = 9.97, p < .01; severity (54.9%), F(2, 97) = 8.74, p < .01; and type of dysarthria (31.7%), F(6, 93) = 4.55, p < .01. For male speakers, the greatest number of subjects was classified with their original severity group, whereas for female speakers, the greatest number of patients was classified with their etiology group. For acoustic variables that are obviously affected by gender (e.g., formant frequencies), both regression and discriminant analyses were conducted separately for male and female speakers.

Table 4.

Classification analysis result for combination of eight variables.

%	All	Male	Female	Contributing factors to classification
Etiology	68.6	56.3	73.7	AR, VID, dB
Type	31.7	39.1	41.0	AR, F0
Severity	54.9	68.3	53.8	F2 slope, F0, VS

Although the focus of this study was on classification accuracy, and not on the individual acoustic variables contributing to the significant discriminant functions, the analysis showed that different sets of acoustic variables contributed to the classification of etiology, type, and severity (see Table 4). Articulation rate, voiceless interval duration, and intensity range contributed to etiology classification, whereas articulation rate and F0 range contributed to type classification. For severity classification, F2 slope, F0 range, and vowel space made significant contributions to the discriminant function.

Classification analysis was performed a final time with only three types of dysarthria (hypokinetic, ataxic, and UUMN; 59 of 107 speakers) because of reservations about unbalanced number of categories (for severity, etiologies, and types of dysarthria), as well as the mixed type of dysarthria that was heterogeneous with respect to dysarthria type. Results showed that the classification rate by types of dysarthria improved when the number of types was reduced from seven to three. However, more subjects were still classified correctly when coded by severity (62.2%) than when coded by type of dysarthria (59.3%) or etiology (56.1%).

Discussion

One goal of this study was to identify a set of acoustic variables that predict speech intelligibility for diverse types of dysarthria and diseases that cause dysarthria and to examine whether different predictors are required for different diseases associated with dysarthria. A second goal was to assess classification accuracy of speakers with dysarthria using acoustic measures as input variables and three different classification variables. The relationship of the findings of this study to these two goals is discussed in the section that follows.

Which Acoustic Variables Predict Speech Intelligibility?

All segmental variables showed significant relationships with intelligibility. This is consistent with the only other study known to us that has compared the impact of segmental and suprasegmental variables on speech intelligibility (de Bodt, Hernández-Díaz, & Van de Heyning, 2002). Among the six variables that had a significant relationship with intelligibility when regression analyses were performed with all speakers pooled, only F2 slope appeared to be significantly regressed on speech intelligibility when analyses were performed within etiology groups (Parkinson’s disease, stroke, traumatic brain injury, and multiple system atrophy). Not only was F2 slope significant for each disease group, but its effect size (i.e., the variance accounted for) was the greatest of all acoustic variables. This is consistent with the results of previous studies that have reported F2 slope as one the most sensitive indices of vocal tract function for speech production—as measured by intelligibility ratings—in neurodegenerative diseases such as amyotrophic lateral sclerosis (R. D. Kent et al., 1989; Weismer et al., 1992) and Parkinson’s disease (Weismer, 1991). F2 slope, therefore, seems to be an indicator of dysarthria severity, as indexed by speech intelligibility scores, for speech motor control deficits in general (i.e., independent of dysarthria type and etiology; see Weismer & Kim, 2010).

In contrast to F2 slope, an exemplar of a “nonuniversal” predictor is articulation rate. Articulation rate was a strong predictor of intelligibility in the pooled analysis but was not a significant predictor of speech intelligibility for speakers with Parkinson’s disease, who were exclusively diagnosed with hypokinetic dysarthria. As reported in several studies, speakers with Parkinson’s disease (or speakers with hypokinetic dysarthria) can have speech rate similar to or faster than (and sometimes slower) that of healthy speakers. This may create a ceiling effect on a possible relationship between articulation rate and intelligibility in this group (Goberman & McMillan, 2005; Weismer, 1984). In other words, there is more margin for the relationship between two factors to vary on the low side of typical articulation rate than to vary on the high side (Turner & Weismer, 1993). Perhaps a reasonable explanation is that faster-than-typical articulation rates do not associate with lowered intelligibility in the same way as do slower-than-typical rates, the latter of which clearly covaried with speech intelligibility in the current study. Alternatively, or in addition to this, because so many studies have found essentially normal rates in Parkinson’s disease, the measure may not be expected to predict variation in speech intelligibility for this group.

F2 slope and articulation rate as predictors of speech severity are also different in another way. Both variables are connected to overall speech severity, but the relationship between articulation rate and speech severity, even excluding speakers with Parkinson’s disease, may be slightly more complicated. Whereas articulation rate decreases with speech severity, as indexed by intelligibility, it may also decrease as a result of speaker compensation to achieve greater intelligibility. In this sense, slowing of articulation rate may reflect influences with opposite effects on speech intelligibility—one effect due to severity, the other due to attempted “correction” for the effects of severity. Compensation for shallow F2 slopes, on the other hand, reflects an attempt to move articulatory behavior in the same direction as lesser severity: Increased transition “speed” is both compensatory and associated with lesser severity (greater intelligibility). Because of these considerations, theories of motor speech disorders and clinical practice may need to treat F2 slope and articulation rate differently. Although there is evidence that speakers with dysarthria can change articulation rate on command (e.g., Van Nuffelen, de Bodt, Vanderwegen, Van de Heyning, & Wuyts, 2010), a similar demonstration of voluntary control over F2 slope for syllables with large vocalic transitions does not exist for speakers with dysarthria.

It is not surprising that the majority of acoustic variables covaried with speech intelligibility scores in this study, especially considering the wide range of speech intelligibility scores for the speakers with dysarthria and the large number of participants in the study. Across speakers, it is clearly a matter of the degree to which an acoustic variable is affected by dysarthria, rather than whether the variable is affected. The fact that all variables are expected to covary with speech intelligibility, especially when intelligibility varies over a wide range, does not necessarily mean the covariation would be the same for different disease groups or dysarthria types. The current analyses were restricted to disease type because similar analyses using dysarthria type would have been partially (and in some cases largely) redundant because of the substantial overlap between disease and dysarthria type in many (but not all) cases (Duffy, 2006). Of course, it is possible that an analysis of acoustic predictors of speech intelligibility within a group of patients having a single disease but multiple types of dysarthria (e.g., head injury or corticobasal degeneration; see Duffy, 2006) would reveal dysarthria-type–specific predictors of intelligibility. The present results cannot answer this question definitively but only suggest that the results would not differ much from those reported here.

Classification of Dysarthria

Since the appearance of the Mayo System, most acoustic studies have sought to discover acoustic correlates that quantify perceptual features of dysarthria as described by Darley et al. (1969a, 1969b) within a single disease or dysarthria type. More specifically, acoustic studies have been conducted to establish a set of acoustic characteristics that describe overall speech characteristics pertaining to a certain disease or dysarthria type, rather than comparing these characteristics across disease or dysarthria types. However, as a substantial amount of speech acoustic data have been collected from persons with various types of dysarthria, it is apparent that a common set of acoustic characteristics is exhibited in multiple types of dysarthria. The presence of acoustic commonalities is not surprising because many of the perceptual dimensions in the Darley et al. studies are shared across dysarthria types, as documented by Darley et al. (1969a) and subsequent studies (Zeplin & Kent, 1996). It is the clusters of dimensions that are thought to distinguish among types of dysarthria, not individual perceptual dimensions. Perceptual clusters are likely to be associated with multiple acoustic indices, and perhaps it is not surprising that a particular acoustic measure would show severity-related variations within several different types of dysarthria. A single measure may be common to different dysarthria types even though its combination with other measures may be a unique feature of a particular dysarthria type.

For example, studies have shown more shallow F2 slopes for transitions of speakers with motor speech disorders, regardless of the etiology and type of dysarthria (Kim et al., 2009). This phenomenon is interesting considering the dissimilarity of the typical underlying neuropathology in diseases such as Parkinson’s disease and stroke. Acoustic similarity across etiologies and types of dysarthria has also been found for vowel space, speaking rate, VOT, and even tone production in Mandarin (evidence reviewed by Weismer, 2006). The current study adds two more acoustic similarities across the four etiology groups in this investigation, including reduced intensity range and the M1 difference between /s/ and /∫/. Accumulating data support the view that different neuropathophysiologies may give rise to similar manifestations at the speech acoustic surface and, by inference, at the level of neuromotor control of speech production (presumably including motor commands and resulting movements of speech mechanism structures). Even with different underlying neuropathologies, there are probably a constrained number of ways in which performance of the speech mechanism can deteriorate when affected by neurological disease. These constraints presumably result in the core set of acoustic features observed across different types of dysarthria, as described previously.

However, the fact that highly trained clinicians or persons with a small amount of intense training can make some reliable distinctions between the different dysarthria types (Zyski & Weisiger, 1987) indicates there is perceptual information (and presumably associated acoustic signal properties) that distinguishes among dysarthria types. The current findings indicate that articulation rate is a feature that makes speakers with Parkinson’s disease (typically associated with hypokinetic dysarthria) distinguishable from other etiologies/types of dysarthria. In addition to (or as part of) faster articulation rates, speakers with Parkinson’s disease also showed shorter voiceless interval durations compared with other etiologies; like speaking rate, voiceless interval duration was not significantly correlated with speech intelligibility. Voiceless interval duration is probably another acoustic manifestation of a relatively normal or fast speaking rate (based on the significant correlation between these two variables, r = .53, p < .0001), and both are apparently unique features of hypokinetic dysarthria associated with Parkinson’s disease.

This observation that only the articulation rate–Parkinson’s disease group pair was distinguishable from other etiologies—and some evidence suggesting that speakers with Parkinson’s disease (and, presumably, hypokinetic dysarthria) are identified with greater accuracy than are speakers with other dysarthria types (Zyski & Weisiger, 1987)—hints at the possibility that listeners may tend to seek only a small number of critical cues for auditory–perceptual judgments of dysarthria type, rather than use combined information from the entire speech signal. Zeplin and Kent (1996) argued that the ratings of certain perceptual features may be more reliable than others. If the acoustic findings of the current study have perceptual salience, one such reliable and salient feature may be a faster-than-normal or normal speaking rate.

In other words, only one or a small number of speech characteristics—such as strain-strangled voice in combination with slow rate and reduced pitch and loudness variability in spastic dysarthria, normal or fast articulation rate in combination with reduced loudness and reduced pitch, and loudness variability in hypokinetic dysarthria—may be the real cues that are pathognomic for a specific type of dysarthria. When these critical features are missing or subdued and only the common speech characteristics are available to listeners, the judgments become less certain and less compatible with the assumed underlying neuropathology–dysarthria type. For example, imprecise consonants is the deviant speech dimension suggested by Darley et al. (1969a, 1969b) to be exhibited in most types of dysarthria. Possibly, imprecise consonants would not cue any particular type of dysarthria unless it could be shown that the nature of consonant imprecision is dysarthria specific.

One of the goals of the present study was to evaluate the classification of speakers with dysarthria using speech acoustic measures and the alternative grouping variables of dysarthria type, disease type (etiology), and speech severity. Regardless of the grouping variable, classification by discriminant function analysis was significant, suggesting that disease type and speech severity may have as much theoretical and clinical relevance to dysarthria as the dysarthria type specified by the Mayo System and its recent refinements. Moreover, when all speakers were pooled for classification analysis, the results (see Tables 3 and 4) indicate that the best classification was obtained when disease type was the grouping variable. The poorest classification performance was obtained when the grouping variable was dysarthria type. Classification accuracy by severity of speech involvement (indexed rather coarsely by speech intelligibility scores) was closer to classification by disease. Taken together, these results seem to suggest not that dysarthria type is an undesirable grouping variable but that it is no more desirable or necessary than the two other classification variables explored here, at least when classification is based on specific acoustic variables. This is an important finding because it may bear on the way in which future experiments on dysarthria choose to group speakers for presumed homogeneity. These results suggest that blocking of participants on dysarthria type, or a requirement that dysarthria type be reported even when disease type is known, may not provide additional insight to interpretation of speech production or perception data obtained from speakers with dysarthria.

If the relatively nonunique classification performance by dysarthria type in the current investigation can be shown to extend to additional predictor variables (including physiological, acoustic, and perceptual variables), the contemporary categorization of motor speech disorders may have to be reconceptualized, at least with respect to the development of a theoretical framework for understanding the disorder and its effect on spoken language and its understanding. On the other hand, there are several reasons to view the implications of the current classification results in a conservative way. First, the identification of dysarthria type was performed by a single person, albeit one with substantial experience in the classification of motor speech disorders using the Mayo System. Perhaps some form of expert panel consideration, with extended listening to speech samples and explicit rubrics concerning the components of particular dysarthria types, would produce a classification of type with greater reliability. Another approach to overcoming the problems with perceptual ratings and classification is to use a computerized system that operates on acoustic metrics. Such a system developed by Carmichael (2007) combines an expert system and a pretrained multilayer perceptron. This system produced good classification results under certain conditions. Second, the results of the different classification analyses reported in Tables 3 and 4 cannot be regarded as fully independent estimates of different classification performance. After all, dysarthria type is often (but not always) linked with disease (etiology) type, and different diseases may be more or less likely to produce greater speech severity in the kinds of utterances from which the acoustic measures were extracted (e.g., persons with Parkinson’s disease are often highly intelligible in prepared utterances even when not so intelligible in spontaneous speech, whereas persons with amyotrophic lateral sclerosis may be equally intelligible in both forms of utterance). Finally, the acoustic measures used in the current investigation were relatively extensive and were selected for their known sensitivity to dysarthria but clearly were not broad enough in scope to rule out the possibility that a critical classifying measure was omitted from the analyses. Perhaps a more extensive set of measures would capture specific speech production phenomena that are the basis for the presumed validity and reliability of the dysarthria classification system originally advanced by Darley et al. (1969a). For example, Liss et al. (2009) recently demonstrated that acoustic metrics of speech rhythm classified persons having four different types of dysarthria with an impressive degree of accuracy (roughly 70%–80% correct classification, depending on the stringency of the classification criteria). The current study did not include the variety of rhythm measures used in the Liss et al. classification exercise, which may explain the better classification performance in the latter study. It must be pointed out, however, that in Liss et al., the dysarthria-type classifications of participants were chosen specifically from among a much larger group of potential participants to have rhythm characteristics typical of the cardinal dysarthria types presumed to be associated with four different diseases (see Liss et al., 2009, p. 1336, and J. M. Liss, personal communication, December 29, 2009).

Participants in the current study were chosen only because they had neuromotor diseases known to produce dysarthria. The classification success in Liss et al. (2009) is, at best, roughly 10% better than the classification by disease in the current study and may decrease for a less selected set of participants. It is also possible that the classification results of Liss et al. were better than those of the present study because they used a normalized version of the PVI (see White & Mattys, 2007), in contrast to the raw PVI measure of the current investigation. Clearly, additional investigations with participants having different disease types, and possibly different dysarthria types, as well as additional measures are required to make progress on the issue of what makes speakers with dysarthria similar to one another and what makes them different from each other.