Effects of Vocal Intensity and Fundamental Frequency on Cepstral Peak Prominence in Patients with Voice Disorders and Vocally Healthy Controls

Abstract

Objective:

Cepstrum-based voice measures, such as smoothed cepstral peak prominence (CPPS), are influenced by voice sound pressure level (SPL) in vocally healthy adults. Since it is unclear if similar effects hold in voice disordered adults and how these interact with natural fundamental frequency (f_o) changes, this study examines voice SPL and f_o effects on CPPS in women with vocal hyperfunction and vocally healthy controls.

Study Design:

Retrospective matched case-control study

Methods:

Fifty-eight women with vocal hyperfunction were individually matched with 58 vocally healthy women for occupation and approximate age. The patient group comprised women exhibiting phonotraumatic vocal hyperfunction associated with vocal fold nodules (n=39) or polyps (n=5), and non-phonotraumatic vocal hyperfunction associated with primary muscle tension dysphonia (n=14). All participants sustained the vowel /a/ at soft, comfortable, and loud loudness conditions. Voice SPL, f_o, and CPPS (dB) were computed from acoustic voice recordings using Praat. The effects of loudness condition, measured voice SPL, and f_o on CPPS were assessed with linear mixed models. Pairwise correlations among voice SPL, f_o, and CPPS were assessed using multiple regression analysis.

Results:

Increasing voice SPL correlated significantly (p<.001) with higher CPPS in both patient (r²=.53) and normative groups (r²=.45). f_o had statistically significant effects on CPPS (p<.001), but with a weak relation for the patient (r²=.02) and control groups (r²⁼.05).

Conclusions:

In women with and without voice disorder, CPPS is highly affected by the individual’s voice SPL in vowel phonation. Future studies could investigate how these effects should be controlled for to improve the diagnostic value of acoustic-based cepstral measures.

1. Introduction

Recent recommendations for a comprehensive standard voice assessment include the cepstrum-based acoustic voice measure Cepstral Peak Prominence (CPP) to objectively describe voice quality in sustained vowels and speech ¹. Cepstral measures estimate the proportion of periodic energy based on a power spectrum ², and thereby indicate the harmonic organisation of an acoustic signal. For clinical voice assessment, cepstrum-based measures have been reported to be reliable indicators of dysphonia, especially in moderately or strongly dysphonic voices ^1-7.

Cepstral voice analysis technique

In a popular diagnostic application of CPP, the computation is derived from the power spectrum of the power spectrum of an acoustic voice signal. CPP indicates the difference (in dB) between the first rahmonic gamnitude and the point of a regression line fitted across the cepstrum that crosses the quefrency of the first rahmonic ^{2, 4}. The Smoothed Cepstral Peak Prominence (CPPS) is a variant of CPP with an additional processing step of smoothing the individual cepstra in the temporal and spectral domains before calculating the peak prominence ^{2, 8}. In regular or Type 1 voice signals, the first cepstral peak (also first rahmonic) corresponds to the fundamental frequency period ^{1, 2, 4, 9}. Acoustic signals with a clear harmonic structure show a more prominent cepstral peak, whereas dysphonic, aperiodic, or breathy voice signals have a reduced cepstral peak. In a variety of studies, CPP has been shown to be correlated with dysphonia severity and was described as an index of the harmonicity of an acoustic signal ^1-6.

In practice, researchers and clinicians often compute CPP using custom algorithms, as provided in commercial software packages such as ADSV (PENTAX Medical Corporation), or freely available software such as PRAAT ¹⁰. For these two software systems, Watts et al. (2017) reported a strong parallel-forms reliability for CPP. The authors concluded that regardless of the language analysed, CPP values could be transformed between programs with relatively small prediction errors ³. Sauder et. al. reported an accuracy of 82% for CPPS in predicting the presence of a voice disorder in connected speech samples when applying the software Praat, and a 75% accuracy for the software ADSV. Despite slight differences in sensitivity and specificity, Sauder et. al. described CPPS as highly predictive of voice disorder regardless of analysis strategy ¹¹. Thus, both CPP and CPPS have a high applicability in clinical practice, but should be communicated with detailed measurement characteristics.

Clinical application of cepstral measures

Cepstral analysis has been recommended in both connected speech samples, such as in recordings of a standard reading passage, and sustained vowels ^{1, 7, 12}. For both contexts, significant differences for CPP and CPPS between healthy and dysphonic voices have been reported ^13-16. Several authors found a similar discriminatory power for the detection of perceptual dysphonia for both sustained vowels and vowels from speech. Based on this, it was argued that the speaking context does not significantly affect CPP measurements in normal and dysphonic voices ^{7, 17,18, 19}. However, a better correlation between perceptual overall dysphonia and breathiness with CPP and CPPS in connected speech than in sustained vowels was also reported ^{6, 14, 20, 21}, or a better discriminatory power of extracted vowels as compared to speech ¹⁵.

Watts et. al. found a higher Cepstral/Spectral Index of Dysphonia (CSID) in older as compared to younger men, but only in tokens from read text and not in vowel phonation. The authors concluded that the clinical voice task might elicit different information about voice function and the underlying vocal physiology ²². Hasanvand et. al. reported significant differences for CPP and CPPS scores in reading tasks in both vocally healthy as compared to voice disordered women and men. However, in sustained vowels there was a significant difference for women only ¹⁶. Further, CPP and CPPS scores were lower in both sustained vowels and reading tasks for dysphonic females as compared to the control group and either group of males. This literature review shows that both cepstral measures CPP and CPPS are relatively reliable and objective assessment tools in distinguishing between dysphonic and normal voices with a high clinical applicability. However, to date it is not yet fully established, which sample (phonation) type is more representative for perceptual dysphonia or voice pathology, and how we best investigate the underlying vocal physiology.

Effects of voice SPL and f_o on cepstral measures

In clinical assessments, patients will produce a comparatively large SPL and f_o range between individuals in response to the usually applied voice task to sustain a vowel “at habitual voice pitch and loudness” ¹. In phonation during both sustained vowels and speech utterances, an association of higher voice SPL with increased f_o has been described ^23-25,26-28. During sustained vowels, phonation with increased voice SPL has been shown to significantly reduce vocal perturbation in individuals with and without vocal pathology. Similarly, increased f_o was associated with a reduction in measures of vocal perturbation ^{26, 29, 30,24, 27}.

However to date, there is only limited evidence available regarding the effects of natural voice SPL and f_o differences on spectral measures in a clinical population. Awan et. al. found increased CPPS in louder phonations in an investigation of vowel and voice SPL effects in 92 vocally healthy women and men between 18 and 30 years of age ³¹. Also, an increase of CPPS in louder voice intensities was reported for sustained vowel and connected speech samples from functionally healthy teachers^{32, 33}. These results indicate that natural variations in voice SPL may confound the interpretation and therefore diagnostic usefulness of cepstral measures in voice assessment. To the best of our knowledge, f_o effects on CPPS in voice patients have not been comprehensively assessed so far. In theoretical studies using synthesized vowel stimuli, Skowronski et al. (2015) showed a decrease in CPP as f_o increased ³⁴. However, natural adaptations in speaking voice f_o require complex adaptations in source properties, which have been shown to affect the acoustic spectrum ³⁵.

Therefore, the main objectives of the present work were to investigate the effects of voice SPL and f_o on the cepstral parameter CPPS computed from sustained vowels produced by individuals with and without diagnosed voice disorders.

2. Methods

Study design and subject characteristics

In a retrospective matched case-control study, data from 116 adult women aged between 18 and 64 years was drawn from a larger project studying vocal hyperfunction using ambulatory voice monitoring ³⁶. Laboratory voice recordings were analyzed from 58 patients diagnosed with phonotraumatic vocal hyperfunction (67.2% with vocal fold nodules, 8.6% with polyps) or non-phonotraumatic vocal hyperfunction (24.1% with muscle tension dysphonia, MTD). Each patient was paired with a vocally healthy control subject who was matched according to sex, approximate age (± 5 years), and occupation/profession. The mean age of the participants with voice disorders was 27.8 years (18–64 years, SD 12.1 years), and the mean age of the matched-control subjects was 27.8 years (18–61 years, SD 11.8 years). There was no statistical difference in age distribution between the groups.

Clinical voice assessments were conducted by a team of laryngologists and speech-language pathologists at the Massachusetts General Hospital Voice Center and included (1) a case history, (2) endoscopic examination of the larynx, (3) aerodynamic and acoustic assessment, (4) the patient-reported Voice-Related Quality of Life (V-RQOL) questionnaire, and (5) clinician-administered Consensus Auditory-Perceptual Evaluation of Voice (CAPE-V) assessment. The normal voice status of all vocally healthy participants was confirmed via interview and a laryngeal stroboscopic examination. All experimental protocols were approved by the institutional review board of Partners HealthCare System at Massachusetts General Hospital.

The level of voice use related to profession was described using the classification scheme of Koufman and Isaacson (1991), modified by do Amaral Catani (2016) ^{37, 38}. This classification scheme is based on voice training and demands and uses four levels. Among the 116 participants, 70 (60%) were rated Level I elite vocal performers (such as singers and actors). 20 (17%) participants were classified as Level II professional voice users (such as teachers) and 16 (14%) as Level III non-vocal professionals (such as social workers). The remaining 10 (9%) were Level IV non-vocal non-professionals, such as administrators and librarians.

Voice recording technique and selection

All participants were asked to sustain a prolonged vowel /a:/ at a comfortable pitch in their typical speaking voice mode at “soft”, “comfortable,” and “loud” voice loudness conditions. Voice recordings were done in a silent room, using a head-mounted microphone integrated in a pneumotachograph mask in an off-axis position at 10 cm distance from the lips (MKE104, Sennheiser, Electronic GmbH, Wennebostel, Germany). The microphone signal was input to a preamplifier (Model 302 Dual Microphone Preamplifier, Symetrix, Inc., Mountlake Terrace, WA, USA), followed by preconditioning electronics (CyberAmp Model 380, Axon Instruments, Inc., Union City, CA, USA) for gain control and anti-alias filtering at a 3 dB cutoff frequency of 8 kHz. The analog signal was digitized at a 20 kHz sampling rate, 16-bit quantization, and ±10 V voltage range (Digidata Model 1440A, Axon Instruments, Inc.).

The acoustic signal from all participants was perceptually examined for instability and visually displayed using the software Praat (version 5.4.1.4) with an oscillogram and “Show intensity” and “Show pulses” settings turned on ¹⁰. Recordings were excluded if they exhibited Type 2 or Type 3 signals, incorrect or unstable f_o and voice SPL recognition in Praat, signal clipping, or a phonation time of less than 1.5 seconds ⁹. Each vowel sample was saved as an individual file. Calibrated voice SPL levels were obtained using the comparison method with a complex tone stimulus of known SPL ³⁹.

Instrumental acoustic analysis and outcome measures

Praat was used to conduct acoustic analysis of the sustained vowel samples using a custom analysis script. Table 1 summarizes how the acoustic outcome measures mean voice SPL, mean f_o, and smoothed cepstral peak prominence (CPPS) were analyzed. To exclude the increased variability of the voice onset and offset phases, only the signal segment from 0.5 second to 1.0 second from voice onset was analysed.

Table 1.

Description of acoustic measures and commands used in Praat

Measure	Unit	Description a, b	Commands in Praat b, c
SPL	dB SPL @10 cm	Sound pressure level of an acoustic signal, calibrated values were determined with comparison method.	1. Sound object: To Intensity (standard settings); 2. Intensity object: Get mean, dB method.
fo	Hz	Main waveform repetition rate (cycles per second), forward cross-correlation method.	1. Sound object: Analyze Periodicity; 2. To Pitch (cc) (standard settings); 3. Pitch object: Get mean.
CPPS	dB	Difference, in gamnitude (amplitude), between the cepstral peak and the corresponding value on the regression line through the cepstrum directly below the peak. The Power Cepstrogram is smoothed by averaging cepstra across time first, and then across quefrency.	1. Sound object: Filter (stop Hann Band). Settings: 0 to 34 Hz, Smoothing 0.1 Hz; 2. Filtered sound object: To Power Cepstrogram (standard settings); 3. Power Cepstrogram object: Get CPPS. Settings: Subtract tilt before smoothing: “no”, Time averaging window: 0.01 s, Quefrency window: 0.001 s, Peak search pitch range: 60–330 Hz, Tolerance: 0.05, Interpolation: Parabolic, Tilt line quefrency range 0.001–0.0 s, Line type: Straight, Fit method: Robust.

dB = decibel, dB SPL @ 10 cm = dB SPL measured at 10 cm mouth-to-microphone distance, Hz = hertz, s = seconds.

^aBaken and Orlikoff (2000)

^bBoersma and Weenink (2018)

^cMaryn and Weenink (2015)

Statistical Analysis

Data were analyzed with the software SPSS^© version 25 (IBM, Armonk, NY). Descriptive analysis comprised of computing the mean (M), standard deviation (SD), 95% confidence interval (CI), and frequency distribution through contingency tables and scatter plot graphics. Normality of the CPPS distributions were assessed by Shapiro-Wilk and Levene’s test. The Shapiro-Wilk test indicated normal distributions for CPPS within both the group with patients [W (167) = .99, p = .18] and healthy controls [W (163) = .99, p = .30]. Furthermore, Levene’s test of equality of variances showed homogeneous distributions of CPPS in both groups [F (1,328) = 1.1, p = .30], and for the categorical voice loudness condition (soft/comfortable/loud) [F (2, 327) = .97, p = .38]. Furthermore, distribution was homogenous for voice SPL [F (1,328) = 0.8, p = .78] and f_o [F (1,328) = 0.1, p = .92] in both groups, and for loudness condition, with SPL [F (2,327) = 2.5, p = .09] and f_o [F (2,327) = 0.25, p = .78].

A multiple regression analysis was applied to investigate the effects of voice SPL (dB SPL@10 cm), f_o (Hz), loudness condition, and presence of a voice disorder (presence and absence) on CPPS. Since repeated measurements tend to be more similar within individuals than across individuals, linear mixed models (LMMs, compound symmetry co-variance type) with loudness condition as repeated measures were applied. Post-hoc analysis was conducted with Bonferroni method to prevent Type 1 errors for multiple comparisons of loudness condition and participant group (confidence interval of 95%, significance level ≤ .05).

3. Results

Effects of SPL, f_o, and loudness condition on CPPS

Table 2 reports group-wide descriptive statistics for mean vocal SPL, mean f_o, and CPPS within the patient and control groups. Signal stability criteria led to the inclusion of an unequal number of recordings per loudness condition in the patient group.

Table 2:

Descriptive statistics for CPPS, SPL, and f_o per loudness condition in the patient and vocally healthy control groups

Acoustic measures	Patient group			Control group
	Soft	comfortable	Loud	Soft	comfortable	loud
SPL (dB SPL)
mean (SD)	79.5 (5.4)	88.0 (4.5)	95.9 (4.3)	81.1 (6.0)	87.7 (5.6)	95.8 (4.7)
CI	78.0–81.0	86.8–89.2	94.8–97.1	79.4–82.7	86.2–89.2	94.5–97.1
fo (Hz)
mean (SD)	248.4 (43.9)	243.3 (41.9)	253.4 (37.8)	244.1 (41.2)	249.2 (36.5)	266.6 (43.6)
CI	236.3–260.5	232.2–254.4	243.4–263.5	232.6–255.6	239.3–258.8	254.7–278.5
CPPS (dB)
mean (SD)	12.3 (2.4)	15.6 (2.0)	18.0 (2.1)	13.3 (2.2)	16 (2.3)	18.0 (2.0)
CI	11.6–12.9	15.1–16.1	17.4–18.5	12.7–13.9	15.4–16.6	17.5–18.5
(N)	53	57	57	52	57	54

Results for mean vocal SPL (dB SPL@10cm), f_o (Hz), and CPPS (dB) with 95% Confidence Interval (CI) per group and loudness condition. N indicates the number of included tokens per loudness conditions in each group.

Table 3 reports the results of the LMM model. When assessed as single factors, both SPL and f_o had highly significant main effects on CPPS in the patient and control groups (p ≤ .001). The loudness condition had a highly significant effect on CPPS in the control group only (p ≤ .001). Further, the interactions between SPL and f_o, f_o and loudness condition, and SPL and loudness condition all had a highly statistically significant effect on CPPS (p ≤ .001). These interaction effects were stronger in the patient group relative to the control group.

Table 3.

Main effects and interactions of voice SPL, f_o, and loudness condition on CPPS within the patient and control groups.

Effect	patient group				control group
	F	estimates (CI)	SER	df	F	estimates (CI)	SER	df
SPL (dB SPL@10 cm)	112.6*	0.31 (0.25–0.36)	0.029	162	34.4*	0.18 (0.13–0.24)	0.026	158
fo (Hz)	81.7*	−0.036 (−0.04–−0.03)	0.003	110	17.6*	−0.02 (−0.03–−0.02)	0.004	152
loudness condition	1.6	−0.85 (−1.9–0.16)	0.51	160	49.9*	−2.50 (−3.4–−0.24)	0.459	151
SPL × fo	27.2*	0.001 (0.001–0.001)	6.01	84	19.7*	0.001 (0.001–0.001)	4.50	160
fo × loudness condition	119.4*	−0.03 (−0.4–−0.001)	0.004	91	78.6*	−0.30 (−0.04–−0.01)	0.005	143
SPL × loudness condition	206.3*	0.26 (0.20–0.32)	0.04	161	116.5*	0.18 (0.13–0.23)	0.03	156

Results of LMM analysis with ANCOVA. Abbreviations: CI = Confidence Interval, SER = Standard Error, df = degree of freedom.

^*p ≤ .001

Figure 1 displays the pairwise relationship between CPPS and SPL, and between CPPS and f_o, within both patient and control groups. In Figure 1A, speakers in both groups exhibit strong correlations for CPPS and SPL (r² =.53 and .45). Figure 1B shows a more spread distribution of f_o values across CPPS results with a weak relation (r² =.02 and .06). On average, there was a 2.2 dB increase in CPPS for a 10 dB increase in SPL. This corresponded to an average CPPS of 13.8 (SD 1.7) for a voice SPL of 80 dB SPL@10cm.

Figure 1.

Scatterplots with linear regression analysis for CPPS versus SPL (dB SPL @10cm) (1A) and CPPS versus f_o (1B) within the patient and control groups. CPPS strongly increased with rising voice SPL, while there was a weak negative correlation between voice fo and CPPS.

Effects of voice disorder on SPL, f_o and CPPS

There was no significant effect of voice disorder on voice SPL and f_o between the patient and control groups for any of the three loudness conditions (soft, comfortable and loud, p>0.5). Considering all loudness conditions together, there was no effect of voice disorder on CPPS [F (1,111) = 1.8, p = .18]. When additionally controlling loudness conditions separately, there were no significant differences between the patient and control groups in comfortable (p = .30) and loud conditions (p = .94). However, in soft phonations CPPS was significantly higher for the control group [t (103) = 2.3, p = .02, Table 2].

4. Discussion

In the present study, CPPS increased with increasing voice intensity in sustained vowels of both female speakers with and without a voice disorder. Although both voice SPL and f_o exhibited statistically significant effects on CPPS, the effect of f_o was far less strong. As with previous studies of perturbation measures, it is hypothesized that the observed strong effect of voice SPL on CPPS will be present in multiparametric indices incorporating CPPS, such as the cepstral-spectral index of dysphonia (CSID) and acoustic voice quality index (AVQI) ^{40, 41}. Based on the results of the current study, it is recommended that voice SPL be controlled for during the clinical assessment of voice and voice quality. Other likely influencing factors such as sex, age, and voice training status should be investigated at a larger scale on patients with a variety of voice disorders.

How relevant are voice SPL and f_o in clinical measurements of CPPS?

In sustained vowel phonations, CPPS increased with increasing voice SPL in women with and without hyperfunctional voice disorders. These effects are similar to results reported for sustained vowel and connected speech samples from functionally healthy teachers with and without laryngeal pathology ^{32, 33}. Thus, patients and vocally healthy individuals with a lower habitual speaking voice SPL will present with reduced periodicity in voicing, which is not related to impairment (i.e. aberrant voice quality). The direct relationship between CPPS and SPL may be a consequence of the stronger harmonic source with increasing voice SPL. Electroglottographic investigations in male singers suggest that glottal closure varies systematically with SPL ⁴². Louder and higher-pitched voicing requires a higher tonus and medial compression of the vocal folds, which may result in improved glottal closure and signal periodicity, as indicated by a higher CPPS ^31,35. Consequently, the relation between an aberrant CPP and physiology still needs to be fully established, since lower (or even higher) CPP is not necessarily associated with the presence of a voice disorder. This puts the application of normative values and thresholds in acoustic voice quality measurements without control of voice SPL in question.

Notably, in the patient group there was significantly lower CPPS in the soft loudness condition. As already described in studies applying the Soft Phonation Index (SPI) and phonation threshold pressure, there may be distinct differences, especially in soft voice production, between speakers with and without voice disorders ^{43, 44}. Thus, CPPS might have indicated increased dysphonia or breathiness in soft phonations of the patient group. As already discussed previously, voice tasks may highlight specific physiological or pathological characteristics ²². However, to date it is not yet fully established, which sample (phonation) type is more representative for perceptual dysphonia or voice pathology, and how we best investigate the underlying vocal physiology. In the present study, f_o was not systematically elicited, and covered only a small part of the human pitch range. Therefore, the present results regarding f_o effects are only valid in relation to the described voice SPL changes in response to the voice task. Again, this emphasizes the potential importance of controlling for voice intensity and the task type during clinical voice assessment.

It has been proposed that voice training may lead to a better production and control of vocal fold tonus, and different acoustic characteristics in vowel phonation, speech and reading tasks ^45-49. In the present study, 60% of participants were vocal performers; i.e., all had undergone professional voice training and were probably subjected to a high vocal demand. Thus, the present results may not be representative for a clinical caseload with mostly untrained voices. Further, men tend to speak louder in the same clinical task under identical measurement conditions ^{26, 28}. As shown by Hasanvand et al, CPP and CPPS scores may be systematically lower in both sustained vowels and reading tasks for females with dysphonia compared to cepstral measures in men. This may be partially due to systematic differences in habitual voice SPL between males and females ¹⁶. Therefore, the present results allow a preliminary estimate of voice SPL and f_o effects in vowel phonation of women. However, further studies including gender, age and trainings status in a variety of voice tasks including speech should be conducted to understand the interrelation between voice SPL, f_o and cepstral measures.

Recommendations for the clinical measurement of CPPS

As implied by the linear regression results of Figure 1A, SPL-related effects may be controlled for by using a correction factor/formula or by reporting CPPS in reference to a standard voice intensity level. However, this should also include an indication of the expected natural variation such as by stating the 95% confidence interval. Also, in our study investigating vowel phonations, we found an average increase of 2.2 dB per 10 dB increase, which was higher than the increase of 1.2 for speech as reported by Phadke et al. The estimated standard value in the present work was 13.8 dB (SD 1.7 dB) for CPPS at 80 dB SPL (@10cm), which was also higher than the previously reported 11.1 dB ³². This may be explained by differences in measurement and analysis techniques, but also by the large voice SPL range covered by our participants. Therefore, a comparison even for SPL-standardized values including Confidence Intervals is only useful when technically identical measurement methods are applied or when data is transformed between programs ³. Further, the voice tasks and the totally produced SPL ranges should be comparable to each other ²².

Another possible approach may be to examine vowels at different loudness conditions such as individually “soft”, “comfortable” and “loud” levels for each speaker to evaluate changes in CPPS over a large SPL span within an individual. Also, we could measure SPL during the first voice examination and ask patients to match the same SPL level in the following assessments. Further, as already discussed for the clinical application of jitter and shimmer measurements, all patients may be asked to produce an identical predefined voice SPL ²⁷. These approaches would minimize the use of a correction formula that has its own uncertainty. However, the voice tasks themselves, and especially controlling for voice SPL, may influence natural voice behaviour, which is presumed to contribute to the voice disorder. Future work should investigate how to best control for technical and procedural confounding factors in the clinical application of CPP and CPPS, to improve the instrumental acoustic assessment of voice quality.

5. Conclusion

In a group of women with and without a voice disorder, the acoustic measure of smoothed cepstral peak prominence (CPPS) was significantly affected by the sound pressure level of the vowel produced. CPPS increased as voice SPL increased. A similar effect has been reported for cepstral measures derived from phonation in vocally healthy individuals and is reproduced here in a group of patients diagnosed with hyperfunctional voice disorders. Since popular voice assessment indices incorporate cepstral measures, it is recommended that the clinical assessment of voice control for variations in voice SPL. Such control may be performed by applying SPL-corrected values for CPPS or by taking care to elicit similar voice SPL levels throughout a patient’s stage of treatment for robust CPPS comparisons. Other potentially confounding factors such as sex, age, and voice training status should be investigated in the future on patients across a larger spectrum of voice disorders.