Reproducibility of Voice Parameters: The Effect of Room Acoustics and Microphones

Abstract

Introduction:

Computer analysis of voice recordings is an integral part of the evaluation and management of voice disorders. In many practices, voice samples are taken in rooms that are not sound attenuated and/or sound-proofed; further, the technology used is rarely consistent. This will likely affect the recordings, and therefore, their analyses.

Objectives.

The objective of this study is to compare various acoustic outcome measures taken from samples recorded in a sound-proofed booth to those recorded in more common clinic environments. Further, the effects from six different commonly used microphones will be compared.

Methods.

Thirty-six speakers were recorded while reading a text and producing sustained vowels in a controlled acoustic environment. The collected samples were reproduced by a Head and Torso Simulator and recorded in three clinical rooms and in a sound booth using six different microphones. Newer measures (e.g. Pitch Strength, CPP, AVQI), as well as more traditional measures (e.g. Jitter, Shimmer, HNR and Spectrum Tilt), were calculated from the samples collected with each microphone and within each room.

Results.

The measures which are more robust to room acoustic differences, background noise, and microphone quality include Jitter and CPPS, followed by Shimmer, AVQI, HNR, Pitch Strength and Spectrum Tilt.

Conclusions.

The effect of room acoustics and background noise on voice parameters appears to be stronger than the type of microphone used for the recording. Consequently, an appropriate acoustical clinical space may be more important than the quality of the microphone.

INTRODUCTION

A voice disorder occurs when voice quality, pitch, and loudness differ or are inappropriate for an individual’s age, gender, cultural background, or geographic location.^1-3 With an occurrence between 20% and 30% in the general population, voice disorders impair an individual’s communicative abilities.^4,5

Evaluation of the dysphonic patient typically involves a combination of laryngeal visualization, perceptual analysis, quality of life surveys, and computerized objective acoustic and aerodynamic measures of the voice signal.⁶ Each of these tools serve a unique purpose, but when used together, they help to best diagnose, manage, and treat a patient with voice disorders. With the increasing emphasis on evidence-based practice in today’s healthcare system, developing and obtaining reproducible outcome measurements is increasingly important.

The evaluation processes mentioned above have been shown to have consistency issues. For example, visualization depends on the tasks the patient performs, the equipment used, and the training and experience of the practitioner. Additionally, perceptual analysis and surveys also have repeatability and reproducibility uncertainties. Computerized acoustic analysis seems to have the potential to be the most objective, but it is also dependent on the speech and voice tasks performed, instructions for the task, equipment used to record, and the acoustic environment.⁷ Many studies involving most traditional acoustic measures of voice and speech have shown significant inconsistency in their ability to quantify outcome measurements, and it is possible that such inconsistency lies within the measurements themselves.⁸ The majority of these measures can only be used for sustained vowel assessment, which may not be an accurate representation of changes in voice quality in running speech.⁹ Moreover, most traditional measures rely on an accurate fundamental frequency estimation, and thus, are only appropriately used for some mildly dysphonic voices and not more severely dysphonic voices.^10,11

With the advancement of more robust objective measures, several acoustic measures have been developed that can be used for running speech and for less periodic signals. These measures include three based on cepstral peak prominence: smooth cepstral peak prominence (CPPS),¹¹ the acoustic voice quality index (AVQI),¹² and more recently, estimated pitch strength (PS).^13-15

Even though these measures can provide a more robust evaluation of the voice, variability in environments may affect the reliability of the results. Previous literature has discussed environmental effects on voice quality and voice assessment.^16-19 However, there is limited research on the effects of the environment on acoustic metrics of voice. In one of these studies, Deliyski et al.²⁰ compared the impact of multiple factors (e.g. microphone and background noise) on the measurement of fundamental frequency (fo), jitter, and shimmer in voice samples from 20 normophonic individuals; the samples were mixed multiple noise types with varying signal-to-noise ratio (SNR) levels. Their results indicated that environmental noise was shown to be the prominent factor that affects voice quality measurements when SNR levels are below 30 dB, while microphone type has significant influences on fo.

In many clinical voice practices (as well as in voice and speech research laboratories), voice samples are taken in wide variety of rooms (ranging in size, background noise level, and reverberation), with different levels of sophistication of equipment. Background noise will likely affect the recordings, and therefore, the analyzed parameters. Previous research showed that variation in background noise at the time of recordings can make difficult to interpret treatment outcomes. An example of this variability is provided in a recent study by Deliyski et al.²¹ This study inspected background noise and its effects on accurate and reliable acoustic voice quality measurements. Their results suggest that the recommended, acceptable, and unacceptable signal-to-noise ratio in the acoustic environment are above 42 dB, above 30 dB, and below 30 dB, respectively. To have an estimation of the signal-to-noise ratio in the clinic environment, it is possible to measure the difference in the noise levels when the patient is recording the speech material and when the patient is quiet. The noise measurements should be performed at the recording microphone position and it can be performed with an accuracy of ± 2 dB using smartphone sound measurement applications (apps).²² In addition, variation in equipment, experience of the clinician, and other characteristics of the environment, such as reverberation time, make it difficult to reliably compare outcomes during a patient’s course of treatment. Furthermore, multi-institutional studies are difficult to perform without measurements that can be consistently and reliably obtained across different centers.

Variability in microphones and other recording equipment is also a factor affecting inconsistencies in acoustic measurement outcomes. Parsa et al. ²³ studied the discriminant effects of different types of microphones on clinically relevant voice parameters. Their results add insight to the differences between microphones when discriminating acoustic parameters such as jitter, shimmer, and harmonics-to-noise ratio (HNR). They found that the type of microphone affects the accuracy of dysphonic voice classification. The 2005 findings of Deliyski et al.²⁴ emphasized the importance of the data acquisition environment (Data Acquisition, DA system and microphone) on the accuracy, sensitivity, and specificity of voice quality measurements. They reported that microphone type is the most important aspect of the DA system. More recently, following a review of technical and voice literature, Švec and Granqvist²⁵ articulated recommendations regarding desirable microphone characteristics that include frequency range, dynamic range, and noise level. Additionally, their study commented on the proximity effect apparent in directional microphones and discussed the placement of microphones in relation to ambient noise levels in the recording environment.

Efforts need to be made to standardize objective analysis of the patients with dysphonia. However, not all facilities can afford to purchase expensive equipment like a sound-proof booth or high-end microphones. Furthermore, it would be difficult to ensure that all clinicians have the same level of expertise or experience. This makes it imperative to better understand the extraneous factors that may impact acoustic measurements of voice. As a part of this goal, the main objective of the current study was to compare how different acoustic environments and microphones impact some of the commonly used acoustic measurements of voice. Acoustic measurements obtained from samples recorded in a sound booth were compared to those recorded in more typical clinic environments using six different types of microphones. We hypothesize that frequency-based measures, such as Jitter and CPPS will be less sensitive to equipment and environmental variables than amplitude-based acoustic measures (such as shimmer and HNR). These amplitude-based measures are more likely to be affected by the filtering effects of the environment’s impulse response compared to frequency-based measures.

EXPERIMENTAL METHOD

The speech materials obtained from 36 speakers served as the test material for evaluating the impact of six different microphones and three clinical environments. This was done by re-playing the original speech materials through a Head and Torso Simulator with Mouth Simulator (HATS, 45BC KEMAR, GRAS, Holte, Denmark). The HATS mouth was previously equalized following the producer’s indication. It was used because it provides acoustic diffraction similar to that encountered around the median human head and torso, both in the proximity and in the far field. The gain of the HATS was set in order to have a speech level of 66 dBA at 1 meter in anechoic conditions. The experimental set-up is shown in Figure 1.

FIG. 1.

Experimental set-up in RoomA.

Speech Materials

Thirty-six college students (19 females and 17 males) between the ages 20 and 25 years old (mean= 22; SE= 0.2) were recorded in a double-walled sound isolation booth (2.5 X 2.75 m and h = 2 m) with a background noise equal to 25 dB(A). All participants were native English speakers with normal hearing (250Hz - 8000 Hz <20dB), and no voice complaints were self-reported the day of the recordings. The speech samples included the first six sentences of the “The Rainbow Passage”²⁶ and three vowels /a/, /i/, and /u/. These were recorded using a class 1 omnidirectional microphone (M2211, NTi Audio, Tigards, OR, USA) connected to a personal computer (PC) running Audacity 2.0.6 (SourceForge, La Jolla, CA) through an audio interface (UH-7000 TASCAM, Teac Corporation, Montebello, CA, USA). The data was saved in an uncompressed format (.wav) at a sampling rate of 44,100 Hz and with 16 bit resolution.

Microphones

The speech samples were recorded with six microphones with differing characteristics and price ranges. The analyzed microphones (described below) were the following:

• M21 - M2211 (NTI Audio, Tigards, OR, USA),

• ECM - ECM8000 (Behringer, Willich, Germany),

• AKG - Perception 120 (AKG, Vienna, Austria),

• CSL - microphone of Computerized Speech Lab (KayPENTAX, Montvale, NJ, USA),

• SHU - Beta 53 (Shure, Niles, IL, USA) and

• LOG - PC Headset 120 (Logitech, Romanel-sur-Morges, Switzerland).

Specifications about the microphones are listed in Table 1. The microphones were amplified with phantom power. The directional microphones were oriented with the main lobe pointing toward the HATS mouth.

Table 1.

Specifications of the 6 microphones analyzed in the experiment.

Mic Id	Microphone	Producer	Polar pattern	Frequency response (Hz)	Sensitivity at 1 kHz	Output impedance (Ohm)	Cost (USD)
M22	M2211	NTI Audio	Omnidirectional	5 - 20000	20 mV/Pa	100	$1,210
ECM	ECM8000	Behringer	Omnidirectional	15 - 20000	11 mV/Pa	600	$60
AKG	Perception 120	AKG	Cardioid	20 - 20000	24 mV/Pa	200	$100
CSL	Computerized Speech Lab Mic	KayPENTAX	Cardioid	55 - 14000	1.3 mV/Pa	150	$100
SHU	Beta 53	Shure	Omnidirectional	20 - 20000	3.16 mV/Pa	136	$350
LOG	PC Headset 120	Logitech	Cardioid	100 - 10000	−39 dBV/Pa	−	$20

Voice parameters evaluation

In the assessment of the voice samples, recently developed measures such as pitch strength, smoothed cepstral peak prominence (CPPS), the acoustic voice quality index (AVQI) and harmonic-to-noise ratio were employed as parameters, as well as more traditional measures such as jitter, shimmer, and spectral slope. All the parameters were calculated with PRAAT version 6.0.13,²⁷ with the exception of pitch strength, which was calculated using a custom script in Matlab R2017a. The script is not publically available, however, it is described in a study by Shrivastav et al. ¹³

Pitch strength is a perceptual attribute that can quantify two sounds with the same pitch, based on the saliency of their pitch sensation.¹³ For example, a pure tone at 500 Hz and a bandpass filtered noise stimulus centered at 500 Hz are perceived to have the same pitch, although the noise stimulus evokes a weaker pitch sensation, and the broader the bandwidth of the filter the weaker pitch sensation. Pitch strength is a tool in the perception of dysphonic voice, as the voices of speakers with dysphonia vary in terms of their pitch strength. Shrivastav et al.¹³ found that low pitch strength was correlated with breathiness and roughness, and that the inclusion of pitch strength in computational models of voice quality may help improve the accuracy of these models.¹³

Measures of smoothed cepstral peak prominence (CPPS) are reliable and strong measures of dysphonia,¹¹ as CPPS is somewhat robust to small errors in fundamental frequency tracking. The CPPS allows for comparison between two testing situations, as it provides the magnitude of the cepstral peak relative to the amplitude of phonation.¹¹

The Acoustic Voice Quality Index (AVQI) is a multivariable model based on acoustic measures that permits the objective assessment of overall dysphonia severity using sustained vowel and continuous speech.¹² The AVQI provides an ecologically valid (i.e., representative of daily voice use) index of dysphonia severity.²⁸ To derive the AVQI, a weighted combination of parameters are modeled in a linear regression formula. The parameters consist of six acoustic measures as follows: time domain (shimmer local, shimmer local dB and harmonics-to-noise ratio (HNR)), frequency domain (general slope of the spectrum and tilt of the regression line through the spectrum) and quefrency domain (smoothed cepstral peak prominences, CPPS).

Harmonics-to-noise ratio (HNR) is a measure that quantifies the relative amount of additive noise in the voice signal.²⁹ It indicates the overall periodicity of the voice signal by quantifying the ratio of the periodic to aperiodic (noise) components.³⁰ When the vocal folds are closed insufficiently, air passes through the glottis, giving rise to turbulence. The resulting frication noise is reflected in a higher noise level in the spectrum. HNR reflects voice quality and is a significant predictor of voice samples that are perceptually rated as rough.²⁹

Jitter is a parameter of frequency variation from cycle to cycle.³⁰ It is commonly seen in dysphonic voice, and its frequency instability is associated with irregular vibration of the vocal folds. The values for Jitter can be measured in different ways, such as absolute, relative, relative average perturbation (rap) and the period perturbation quotient (ppq5). These Jitter outcomes compare the average absolute difference of a period with the surrounding periodicity, or the periodicity of the entire waveform.

Shimmer is a time-based parameter expressed in decibels. It describes the amplitude variation of a waveform, measuring the instability of the amplitude. Shimmer is expressed as the variability of the peak-to-peak amplitude in decibels,³⁰ and it is a useful parameter in checking the micro-variations of the amplitude in adjacent cycles. Clinically, shimmer represents amplitude stability as it changes with the reduction of glottal resistance and mass; it is correlated with the presence of breathiness and noise emission.

Spectral Tilt (or slope) is a descriptive parameter for the overall perception of voice quality. It measures how the amplitudes of successive components decrease with increasing harmonic number.³¹ In this study, the slope was calculated as the direct linear fit of the spectrum between 50-5000 Hz at the analysis window size (10 msec).

Room acoustic parameters

The voice recordings were performed in 4 different environments: a sound-treated booth (RoomA) which served as a baseline, and three different clinical spaces (Room1, Room2 and Room3) which served as the test conditions for the purpose of this study.

The recordings performed in RoomA were used as “true values” in the comparisons because this room had a low level of background noise and reverberation time. Room acoustic parameters were measured in an unoccupied state from the impulse responses (IRs) generated by a balloon pop.³² The volume of this room was 38 m³, the mid-frequency reverberation time, T₃₀, was 0.05 s and the A-weighted equivalent level of the background noise was 20.5 dB(A). The volume of Room1 was 22 m³, the mid-frequency reverberation time, T₃₀, was 0.31 s (s.d. 0.04) and the A-weighted equivalent level of the background noise was 47.5 dB(A). The volume of Room2 was 25 m³, the mid-frequency reverberation time, T₃₀, was 0.26 s (s.d. 0.03) and the A-weighted equivalent level of the background noise was 47.4 dB(A). The volume of Room3 was 32 m³, the mid-frequency reverberation time, T₃₀, was 0.43 s (s.d. 0.04) and the A-weighted equivalent level of the background noise was 49.0 dB(A). The background noise and the reverberation time were measured with the microphone M2211 connected to the XL2 Audio and Acoustic Analyzer (NTI Audio, Tigards, OR, USA). The trend over the octave bands for the reverberation time and the background noise levels of the four environments are reported in Table 2.

Table 2.

Reverberation time (T₃₀) and background noise measured in the 4 rooms per octave band.

Parameter (s.d.)	Room	125 Hz	250 Hz	500 Hz	1000 Hz	2000 Hz	4000 Hz	8000 Hz
T30 (s)	RoomA	0.05	0.04	0.04	0.04	0.04	0.04	0.04
T30 (s)	Room1	0.31	0.35	0.32	0.29	0.29	0.29	0.28
T30 (s)	Room2	0.27	0.34	0.28	0.23	0.24	0.25	0.23
T30 (s)	Room3	0.35	0.40	0.47	0.40	0.39	0.39	0.34
Background noise Spectrum (dBA)	RoomA	−5.0	0.8	6.5	10.6	13.5	14.9	13.9
Background noise Spectrum (dBA)	Room1	42.1	43.7	40.1	35.0	29.4	29.0	24.4
Background noise Spectrum (dBA)	Room2	39.9	43.9	37.2	35.8	37.6	35.4	27.4
Background noise Spectrum (dBA)	Room3	40.6	45.7	41.3	39.6	36.1	28.8	21.7

Statistical method

Statistical analysis was conducted using R version 3.1.2.³³ Linear mixed models (LMEs) fit by restricted maximum likelihood (REML) were built using lme4,³⁴ lmerTest³⁵ and multcomp³⁶ packages.

For each voice parameter, two sets of LMEs were fit. A first set considering only the results in RoomA to evaluate the difference of each microphone from the best available (M21) without the possible confounding factors introduced by the different room acoustics. Subject id and type of task (Speech or Vowel) were chosen as random effects terms to avoid their contributions to the variance. The type of microphone was chosen as a fixed effect. Tukey’s post-hoc pair-wise comparisons³⁷ with single-step correction were performed to examine the differences between all levels of the fixed factors of interest. Among them only the differences of all the microphones from the M21 were reported.

A second set of LMEs was used to assess the effect of room acoustics on the voice parameters. In this case, all the data collected were inserted in the models. Subject id and type of task (Speech or Vowel) and type of microphone were chosen as random effects terms to avoid their contributions to the variance.

The model output included estimates of fixed effects coefficients, standard error associated with the estimate, degrees of freedom, df, the test statistic, t and the p value. The Satterthwaite method³⁷ was used to approximate degrees of freedom and calculate p values, because it produces more conservative p values compared to other methods such as the likelihood ratio test.

The concept of Compatibility Index³⁸ was adopted for the analysis of compatibility between two sets of data obtained in different conditions (compatibility between microphones and between rooms). This test is used to determine whether the difference in the compared models is due to an effective difference between the evaluated phenomena, or to systematic effects, rather than to random effects. The Compatibility Index is calculated as the ratio between the absolute value of the difference between the mean of the two samples and the relative expanded uncertainty of the difference,³⁹ according to the following formula:

$$Compatibility Index=\frac{\left|m_1-m_2\right|}{\text{U}}=\frac{\left|m_1-m_2\right|}{\text{k}\sqrt{s_1^2+s_2^2}}$$ 1

where$m_1$ and$m_2$ represent the average values of the two samples,$s_1$ and $s_2$ represent the standard deviations of the two samples and k is the coverage factor, calculated as the Student-t value for a conventional risk of error α of 5% and a number of degrees of freedom corresponding to $n-2$, where n is the number of samples used. This analysis can be considered a particular kind of hypothesis test. If the Compatibility Index value is higher than unity, the difference between the mean of the two samples, m₁ and m₂, is higher than its uncertainty. Therefore, the difference is not merely due to random effects, and the two results can be considered incompatible. Moreover, the higher the index value, the stronger the incompatibility.

RESULTS

For each voice parameter two models were fitted. Using the data collected in the sound booth, a first linear mixed effect model was fitted with the voice parameter as response variable and microphone as covariate. The random effects terms were subject id and type of task (speech or vowel). The reference level for the microphone was M21. The output of the models for each parameter is reported in Table 3.

Table 3.

LME models fit by REML for the response variables Pitch Strength, Smoothed Cepstral Peak Prominence, Acoustic Voice Quality Index, Jitter, Shimmer, Harmonic to Noise Ratio, and Spectrum Tilt including as fixed factor the type of microphone and as random effects term: subject and type of task (with the exception of AVQI). Reference level is the M21 microphone.

Fixed factors	Estimate	Std. Error	df	t	p	Fixed factors	Estimate	Std. Error	df	t	p
Pitch Strength						Smoothed Cepstral Peak Prominence
(Intercept)	0.380	0.079	2	4.796	0.039*	(Intercept)	21.340	2.434	2	8.767	0.012*
ECM	−0.054	0.005	827	−10.311	<0.001***	ECM	−0.684	0.293	827	−2.337	0.020*
AKG	0.016	0.005	827	3.086	0.002**	AKG	−0.110	0.293	827	−0.376	0.707
CSL	−0.031	0.005	827	−5.902	<0.001***	CSL	−0.310	0.293	827	−1.060	0.289
SHU	−0.059	0.005	827	−11.275	<0.001***	SHU	−0.960	0.293	827	−3.281	0.001**
LOG	−0.139	0.005	827	−26.582	<0.001***	LOG	−0.231	0.293	827	−0.789	0.431
Acoustic Voice Quality Index						Jitter
(Intercept)	3.684	0.156	40	23.562	<0.001***	(Intercept)	0.0039	0.0006	35	6.443	<0.001***
ECM	0.631	0.053	180	11.844	<0.001***	ECM	0.0001	0.0004	612	0.163	0.871
AKG	−0.207	0.053	180	−3.878	<0.001***	AKG	0.0001	0.0004	612	0.263	0.792
CSL	0.320	0.053	180	6.006	<0.001***	CSL	0.0001	0.0004	612	0.260	0.795
SHU	0.728	0.053	180	13.661	<0.001***	SHU	0.0003	0.0004	612	0.831	0.406
LOG	1.799	0.053	180	33.744	<0.001***	LOG	0.0010	0.0004	612	2.798	0.005**
Shimmer						Harmonic to Noise Ratio
(Intercept)	0.028	0.004	24	6.582	<0.001***	(Intercept)	26.168	0.834	11	31.383	<0.001***
ECM	0.009	0.003	610	2.771	0.006**	ECM	−2.070	0.362	610	−5.711	<0.001***
AKG	−0.002	0.003	610	−0.729	0.466	AKG	0.084	0.362	610	0.232	0.817
CSL	−0.003	0.003	610	−0.826	0.409	CSL	−1.569	0.362	610	−4.328	<0.001***
SHU	0.017	0.003	610	5.326	<0.001***	SHU	−3.239	0.362	610	−8.938	<0.001***
LOG	0.076	0.003	610	23.486	<0.001***	LOG	−11.328	0.362	610	−31.256	<0.001***
Spectrum Tilt
(Intercept)	−0.732	0.209	6	−3.504	0.014*
ECM	−0.509	0.049	610	−10.303	<0.001***
AKG	−0.490	0.049	610	−9.918	<0.001***
CSL	1.703	0.049	610	34.493	<0.001***
SHU	−1.360	0.049	610	−27.534	<0.001***
LOG	−0.254	0.049	610	−5.151	<0.001***

Signif. Codes:

^’***’<0.001

^’**’<0.01

^’*’<0.05

’.’<0.1

A second linear mixed effect model was fitted with the voice parameter as response variable and room as covariate. In the analysis of the recording environment, all the data collected was used. The random effects terms were subject id, type of task (speech or vowel) and the type of microphone. The reference level for the room was RoomA (an anechoic sound booth). The output of the models for each parameter is reported in Table 4.

Table 4.

LME models fit by REML for the response variables Pitch Strength, Smoothed Cepstral Peak Prominence, Acoustic Voice Quality Index, Jitter, Shimmer, Harmonic to Noise Ratio, and Spectrum Tilt including as fixed factor the Room and as random effects term: subject, microphone and type of task (with the exception of AVQI). Reference level is the RoomA.

Fixed factors	Estimate	Std. Error	df	t	p	Fixed factors	Estimate	Std. Error	df	t	p
Pitch Strength						Smoothed Cepstral Peak Prominence
(Intercept)	0.347	0.067	2	5.149	0.023*	(Intercept)	21.235	2.076	2	10.230	0.007**
Room1	−0.104	0.003	3414	−31.724	<0.001***	Room1	−1.168	0.122	3414	−9.581	<0.001***
Room2	−0.073	0.003	3414	−22.142	<0.001***	Room2	−0.835	0.122	3414	−6.852	<0.001***
Room3	−0.130	0.003	3414	−39.573	<0.001***	Room3	−1.894	0.122	3414	−15.541	<0.001***
Acoustic Voice Quality Index						Jitter
(Intercept)	4.230	0.317	10	13.339	<0.001***	(Intercept)	0.004	0.004	8	1.094	0.305
Room1	1.593	0.068	823	23.333	<0.001***	Room1	0.006	0.001	2549	8.149	<0.001***
Room2	1.211	0.068	823	17.740	<0.001***	Room2	0.005	0.001	2549	6.308	<0.001***
Room3	1.797	0.068	823	26.329	<0.001***	Room3	0.012	0.001	2549	15.084	<0.001***
Shimmer						Harmonic to Noise Ratio
(Intercept)	0.044	0.022	7	2.064	0.077*	(Intercept)	23.147	2.035	9	11.376	<0.001***
Room1	0.073	0.003	2549	27.675	<0.001***	Room1	−9.278	0.298	2549	−31.152	<0.001***
Room2	0.059	0.003	2549	22.572	<0.001***	Room2	−7.447	0.298	2549	−25.002	<0.001***
Room3	0.085	0.003	2549	32.512	<0.001***	Room3	−10.405	0.298	2549	−34.935	<0.001***
Spectrum Tilt
(Intercept)	−0.884	0.481	8	−1.839	0.105
Room1	−2.064	0.048	2549	−42.675	<0.001***
Room2	−1.805	0.048	2549	−37.319	<0.001***
Room3	−2.227	0.048	2549	−46.063	<0.001***

Signif. Codes:

^’***’<0.001

^’**’<0.01

^’*’<0.05

’.’<0.1

Pitch Strength

Figure 2 shows the mean and the standard error of Ps measured with the six microphones in the sound booth for the speech and the vowel tasks. The estimate of standard deviations for random effects were 0.04 for subject id and 0.11 for task, while the residual standard deviation was 0.04. Among all the microphones, the AKG measured the closest values to the M21 (0.01 higher). All the other microphones underestimated the values of Ps. The average differences were −0.03 for CSL, −0.05 for ECM, - 0.06 for SHU and −0.14 for LOG.

FIG. 2.

Mean and the standard error of pitch strength (Ps) measured with the six microphones in the sound booth for the speech and the vowel tasks.

Figure 3 shows the mean and the standard error of Ps measured in the four rooms for the speech and the vowel tasks. The estimate of standard deviations for random effects were 0.04 for subject id, 0.09 for task and 0.05 for the microphone, while the residual standard deviation was 0.07. In all the rooms the values measured were lower than the ones measured in RoomA. The differences were larger in more reverberant rooms. The average differences from RoomA were −0.10 for Room1, −0.07 for Room2 and −0.13 for Room3.

FIG. 3.

Mean and the standard error of pitch strength (Ps) measured in the four rooms for the speech and the vowel tasks.

Smoothed Cepstral Peak Prominence (CPPS)

Figure 4 shows the mean and the standard error of CPPS measured with the six microphones in the sound booth for the speech and the vowel tasks. The estimate of standard deviations for random effects were 2.48 for subject id and 1.49 for task, while the residual standard deviation was 3.41. All the microphones underestimated the values of CPPS. The difference between the M21 and AKG, CSL and LOG were non-statistically significant and lower than 0.31, while ECM and SHU measured values significantly lower than the ones measured by M21.

FIG. 4.

Mean and the standard error of Smoothed Cepstral Peak Prominence (CPPS) measured with the six microphones in the sound booth for the speech and the vowel tasks.

Figure 5 shows the mean and the standard error of CPPS measured in the four rooms for the speech and the vowel tasks. The estimate of standard deviations for random effects were 1.55 for subject id, 2.87 for task and 0.78 for the microphone, while the residual standard deviation was 2.53. In all the rooms the values measured were lower than the ones measures in RoomA. The differences were larger in more reverberate rooms. The average differences from RoomA were −1.17 for Room1, −0.83 for Room2 and −1.89 for Room3.

FIG. 5.

Mean and the standard error of Smoothed Cepstral Peak Prominence (CPPS) measured in the four rooms for the speech and the vowel tasks.

Acoustic Voice Quality Index (AVQI)

Figure 6 shows the mean and the standard error of AVQI measured with the six microphones in the sound booth. The estimate of standard deviation for random effect “subject id” was 0.91, while the residual standard deviation was 0.23. Among all the microphones, the AKG measured the closest values to the M21 (0.21 lower). All the other microphones overestimated the values of AVQI. The average differences were 0.32 for CSL, 0.63 for ECM, 0.73 for SHU and 1.80 for LOG.

FIG. 6.

Mean and the standard error of Acoustic Voice Quality Index (AVQI) measured with the six microphones in the sound booth for the speech and the vowel tasks.

Figure 7 shows the mean and the standard error of AVQI measured in the four rooms. The estimate of standard deviations for random effects were 0.85 for subject id and 0.68 for task, while the residual standard deviation was 0.71. In all the rooms the values measured were higher than the ones measures in RoomA. The differences were larger in more reverberate rooms. The average differences from RoomA were 1.59 for Room1, 1.21 for Room2 and 1.79 for Room3.

FIG. 7.

Mean and the standard error of Acoustic Voice Quality Index (AVQI) measured in the four rooms for the speech and the vowel tasks.

Jitter

Figure 8 shows the mean and the standard error of jitter measured with the six microphones in the sound booth for the speech and the vowel tasks. The estimate of standard deviations for random effects were 0.003 for subject id and 0.0004 for task, while the residual standard deviation was 0.003. All the microphones measured similar values with the exception of LOG, which overestimated the parameter (0.001 higher than M21).

FIG. 8.

Mean and the standard error of Jitter measured with the six microphones in the sound booth for the speech and the vowel tasks.

Figure 9 shows the mean and the standard error of jitter measured in the four rooms for the speech and the vowel tasks. The estimate of standard deviations for random effects were 0.004 for subject id, 0.004 for task and 0.007 for the microphone, while the residual standard deviation was 0.014. In all the rooms the values measured were lower than the ones measures in RoomA. The differences were larger in more reverberate rooms. The average differences from RoomA were 0.006 for Room1, 0.005 for Room2 and 0.12 for Room3.

FIG. 9.

Mean and the standard error of Jitter measured in the four rooms for the speech and the vowel tasks.

Shimmer

Figure 10 shows the mean and the standard error of shimmer measured with the six microphones in the sound booth for the speech and the vowel tasks. The estimate of standard deviations for random effects were 0.017 for subject id and 0.003 for task, while the residual standard deviation was 0.024. Among all the microphones, the AKG measured the closest values to the M21 (0.01 higher). The microphones AKG and CSL measured similar values compared to the one measured by M21.The average differences for other microphones were statistically significant and they were 0.009 for ECM, 0.017 for SHU and 0.076 for LOG.

FIG. 10.

Mean and the standard error of Shimmer measured with the six microphones in the sound booth for the speech and the vowel tasks.

Figure 11 shows the mean and the standard error of shimmer measured in the four rooms for the speech and the vowel tasks. The estimate of standard deviations for random effects were 0.02 for subject id, 0.03 for task and 0.04 for the microphone, while the residual standard deviation was 0.05. In all the rooms the values measured were lower than the ones measures in RoomA. The differences were larger in more reverberate rooms. The average differences from RoomA were 0.073 for Room1, 0.059 for Room2 and 0.085 for Room3.

FIG. 11.

Mean and the standard error of Shimmer measured in the four rooms for the speech and the vowel tasks.

Harmonic to Noise Ratio (HNR)

Figure 12 shows the mean and the standard error of HNR measured with the six microphones in the sound booth for the speech and the vowel tasks. The estimate of standard deviations for random effects were 3.28 for subject id and 0.99 for task, while the residual standard deviation was 2.66. Among all the microphones, the AKG measured the closest values to the M21 (0.08 higher). All the other microphones underestimated the values of HNR. The average differences were −1.57 for CSL, −2.07 for ECM, −3.24 for SHU and −11.33 for LOG.

FIG. 12.

Mean and the standard error of Harmonic to Noise Ratio (HNR) measured with the six microphones in the sound booth for the speech and the vowel tasks.

Figure 13 shows the mean and the standard error of HNR measured in the four rooms for the speech and the vowel tasks. The estimate of standard deviations for random effects were 2.93 for subject id, 4.20 for task and 1.65 for the microphone, while the residual standard deviation was 5.36. In all the rooms the values measured were lower than the ones measures in RoomA. The differences were larger in more reverberate rooms. The average differences from RoomA were −9.28 for Room1, −7.45 for Room2 and −10.40 for Room3.

FIG. 13.

Mean and the standard error of Harmonic to Noise Ratio (HNR) measured in the four rooms for the speech and the vowel tasks.

Spectrum Tilt

Figure 14 shows the mean and the standard error of Ps measured with the six microphones in the sound booth for the speech and the vowel tasks. The estimate of standard deviations for random effects were 0.64 for subject id and 0.31 for task, while the residual standard deviation was 0.36. Among all the microphones, the CSL measured the most different values compared to the one measured by the M21 (1.70 higher). All the other microphones underestimated the values of Spectrum Tilt. The average differences were −0.51 for ECM, −0.49 for AKG, - 1.36 for SHU and −0.25 for LOG.

FIG. 14.

Mean and the standard error of Spectrum Tilt measured with the six microphones in the sound booth for the speech and the vowel tasks.

Figure 15 shows the mean and the standard error of Spectrum Tilt measured in the four rooms for the speech and the vowel tasks. The estimate of standard deviations for random effects were 0.54 for subject id, 1.09 for task and 0.26 for the microphone, while the residual standard deviation was 0.87. In all the rooms the values measured were lower than the ones measured in RoomA. The differences were larger in more reverberate rooms. The average differences from RoomA were −2.06 for Room1, −1.80 for Room2 and −2.23 for Room3.

FIG. 15.

Mean and the standard error of Spectrum Tilt measured in the four rooms for the speech and the vowel tasks.

Compatibility of the results among microphones and rooms

In order to test the compatibility of the results for the different parameters among the microphones in perfect acoustic conditions (RoomA), the compatibility index was calculated following Eq (1). With regard to microphone performance across all parameters, the AKG microphone gave the closest results compared to the M21, followed by the ECM, the CSL, SHU and the LOG. Among the parameters, the most independent to variations in microphones were the Jitter and the CPPS, followed by Shimmer, HNR, Ps, AVQI and Spectrum tilt. Figure 16 shows these results.

FIG. 16.

Compatibility indices of the 7 parameters (Ps, CPPS, AVQI, Jitter, Shimmer, HNR and Spectrum tilt) performed with the different microphones (AKG, CSL, ECM, SHU, LOG) compared to the M211 in RoomA.

Examining the effect of the three room characteristics, Room2 gave the closest results compared to RoomA, followed by Room1 and Room3. Among the parameters, the most independent of the varying acoustics of the room were Jitter and CPPS, followed by AVQI, Shimmer, HNR, Ps and Spectrum tilt. Figure 17 shows these results.

FIG. 17.

Compatibility indices of the 7 parameters (Ps, CPPS, AVQI, Jitter, Shimmer, HNR and Spectrum tilt) performed in the three different rooms (Room1, Room2, Room3) compared to RoomA.

DISCUSSION

In the clinical evaluation of dysphonic voices, and in the quantification of voice improvement during speech therapy, it is extremely important to estimate the reliability of acoustical voice parameters. However, these parameters often exhibit poor reproducibility between studies. Among the various sources of uncertainty in the evaluation of voice parameters, the most significant factors include the acoustics of the recording environment and the microphone used during the recordings.

Considering the heterogeneity of room acoustic features in clinical environments, of the equipment utilized for the measurements, and of the acoustic parameters commonly used in clinics, the aim of this study was to compare a number of acoustic outcome measures gathered from samples recorded in a sound booth to those recorded in more common clinic environments, as well as with different microphones.

This study uncovered the relationship between microphone type and voice quality measurements, as Figure 16 demonstrates. Comparing five commercial and professional grade microphones to a reference omnidirectional microphone (M21), comparison data suggests which microphone is the most accurate for voice quality measurements in ideal acoustic conditions and in typical clinical environments. Parsa et al.²³ employed a similar methodology to discover the effects of microphone type on acoustic measures of voice. Their study tested three unidirectional (cardioid) microphones and one measurement-quality, omnidirectional reference microphone. They examined the discrimination of various acoustic parameters including jitter, shimmer, and HNR. Their results concluded that all microphone characteristics alter the voice signal during the recording process, thus possibly reducing the discriminatory abilities of the acoustic parameters. They extended these results, examining how the choice of microphone affects key voice parameters and investigated how one might compensate for such microphone effects through filtering or by including additional parameters in the decision process. The results of the present study confirm and expand upon Parsa et al.’s²³ conclusions. While the Parsa et al. study did not specify the microphone brand, their use of three cardioid microphones and one omnidirectional reference microphone was comparable to the present study’s use of three omnidirectional and three cardioid microphones, and our findings are comparable as well. Specifically, our studies agree in finding variation with the HNR measurement among the microphones, as Parsa et al. cited significant HNR variation in all microphones expect the omnidirectional. In examining the discriminatory acoustic differences between five microphones of varying levels, the present study compiled compatibility data to reveal that the AKG microphone gave the closest results compared to the M21, followed by the ECM, the CSL, SHU and the LOG. The average compatibility indexes for the aforementioned microphones were 1.3 for AKG, 3.2 for ECM, 3.8 for CSL, 5.1 for SHU and 9.1 for LOG. Regarding the interaction between acoustic voice quality parameters and microphone quality, the most robust parameters among the different microphones were Jitter and CPPS, followed by Shimmer, HNR, Ps, AVQI and Spectrum tilt. The average compatibility indexes for the parameters were 0.4 for Jitter, 0.8 for CPPS, 3.6 for Shimmer, 5.1 for HNR, 6.0 for Ps, 7.0 for AVQI and 8.8 for Spectrum tilt.

Building upon the findings of Parsa et al.²³, Deliyski et al.²⁴ inspected the voice sample data acquisition environment as a whole. Their study specifically focused on the importance of hardware selection when gathering acoustic voice quality measurements. Employing three software systems and six microphones to record and analyze voice samples, their study assessed several parameters including fundamental frequency, jitter, and shimmer. The results concluded that microphone type is the most important factor in the data acquisition (DA) environment’s effect on voice quality measurements. Additionally, the study demonstrates assertively that DA systems that do not accept head-mounted condenser microphones with balanced output should not be used for acoustic assessment of voice quality.²⁷ The present study investigated a variety of microphones including two head-mounted microphones (SHU and LOG). However, the compatibility data suggests that the head-mounted microphones performed poorly in comparison to the other microphones that were tested. These results are consistent with the findings of Švec and Granqvist²⁵ who noted that mouth-to-microphone distances smaller than 20 cm may cause problems for accurate spectral measurements of voice and speech.

Švec and Granqvist²⁵ also examined recommended environmental noise levels, and their association with microphone distance. Their review concluded that room noise levels should be lower than 40 dBA when the microphone is put on a stand and placed at 30 cm distance from the lips. Moreover, they cited Dejonckere et al.⁴⁰ when describing head-mounted microphones, stating that when the microphone is head-mounted at 10 cm distance from the lips “recordings should be made ideally in a sound-treated room, but a quiet room with ambient noise < 50 dB is acceptable.”⁴⁰ These recommendations shed light on the discrepancies between the present study and the findings of Deliyski et al.²³ Specifically, it is important to note Dejonckere’s⁴⁰ findings when considering the distinction between recommended noise level for microphones on a stand versus those that are head-mounted. In the present study, Rooms 1, 2 and 3 conveyed A-weighted equivalent levels of background noise that measure well above 40 dBA. This level of background noise has clinical significance, as the three test rooms are examples of common clinical testing environments. Therefore, as evidenced by Dejonckere et al.⁴⁰ and the results of the present study, clinicians should consider both microphone type and environmental noise in voice measurement rooms. The results of this study are showing that the effect of room acoustics on the repeatability of the voice parameters is stronger than the effect of the microphone. This means that a proper acoustic design of the clinical space, aimed to minimize the reverberation time and maximize the noise insulation, is necessary.

The range of recommendations in studies that compare microphone type is evidence that microphone placement is a salient issue which beckons future investigative efforts in order to provide well-founded guidelines for voice quality measurements.

Regarding room acoustics, this study uncovered the relationship between reverberation time, background noise and voice quality measurements. Under the hypothesis that some acoustic measures will be less sensitive to equipment and environmental variables than others, the results confirm that the acoustic parameters which were tested vary according to the acoustics of the environment. The measures which are more notably robust (less sensitive) to room acoustic differences, background noise, and microphone quality include Jitter and CPPS. Figures 16 and 17 illustrate the variance in the acoustic parameters of the voice samples with five microphones compared to a reference level (M21), as well as with the three clinical recording environments compared to a reference level (Room A).

In examining the discriminatory acoustic differences, the present study compiled compatibility data to reveal that parameters recorded in Room2 were the most similar to those recorded in the reference sound booth. The average compatibility indexes for the three rooms were 9.7 for Room2, 12.3 for Room1 and 14.8 for Room3. Regarding the interaction between acoustic voice quality parameters and room acoustics, the most robust parameters among the different rooms were Jitter and CPPS, followed by AVQI, Shimmer, HNR, Ps, and Spectrum tilt. The average compatibility indexes for the parameters were 3.8 for Jitter, 5.3 for CPPS, 11.3 for AVQI, 12.1 for Shimmer, 15.2 for HNR, 17.1 for Ps and 21.2 for Spectrum tilt.

The significance of the testing confirms the results from Deliyski et al.²⁰ (Regression Tree Approach). In reporting the impact of environmental acoustics, environmental noise (including noise level and noise type) was shown to be the most prominent factor that affected voice quality measurements.²⁴ Table 2 provides the reverberation time (T30) and background noise measured in the four rooms per octave band is provided. The A-weighted equivalent level of background noise in the anechoic room used for “true values” (RoomA) was 20.5 dB(A). Whereas in Rooms 1, 2 and 3, the A-weighted equivalent level of the background noise was 47.5 dB(A), 47.4 dB(A), and 49.0 dB(A) respectively. Based on the compatibility index in Figure 17, the effects of background noise and reverberation time are demonstrated. Room2 displays the lowest background noise and reverberation, and, in doing so, it shows the highest compatibility. In comparison to the values gathered from the other rooms, it is clear that an environment with the lowest possible background noise and reverberation is the best measurement environment when considering acoustic voice quality parameters.

CONCLUSIONS

The results of this study showed that the measurements closest to the best estimate of measures (the results from the M21 in RoomA) were recorded in the room with lower reverberation (Room2) and with the microphone AKG. Among the parameters, the most independent by the acoustics of the room were the Jitter and the CPPS.

In conclusion, comparing the compatibility indexes obtained in different rooms with those obtained by different microphones in the reference sound booth, a notable distinction shows that the effect of room is higher than the incompatibility associated with the effect of microphones. Therefore, it is extremely important to consider an appropriate acoustical design of clinical spaces in order to perform reliable acoustic voice parameters measurements.

One limitation of this study is the number of microphones tested. Although microphones of different types were used in this study, unresolved information regarding mouth-to-microphone distance remains. By analyzing a wider range of microphones and adjusting their orientation, specified standards for voice measurements can be gathered. Additionally, a limited variety of room types was employed in this study. A larger number of testing rooms would provide a more accurate picture of the different environments that are currently used in voice measurement. This study was limited by the lack of dysphonic voices in the sample. By testing speakers with normal voice quality, information produced by dysphonic voices is absent in our results. The last point that needs to be further studied is the clinical significance of the differences observed. Under the statistical point of view, it can be stated that room acoustics and type of microphone are affecting the voice parameters. However, are the differences observed significant for clinicians? This is a complicated question to answer because it depends on the speaker. If a speaker with normal voice quality is close to the cut-off value for a given parameter, he/she may be wrongly labeled as having dysphonic voice quality due to the effect of different room acoustics and/or type of microphone. In the future, a project that includes a wider range of microphones, testing environments, and voice quality within the sample population will produce more substantial results commenting on the effects of room acoustics and microphones on voice parameters. Additional studies are planned that include the use of indoor acoustic simulation software. This will allow us to have a larger variety of reverberation and background noise levels, which could help in the identification of cut-off values for optimal noise and reverberation in clinical spaces.