The Effect of Noise on Computer-Aided Measures of Voice: A Comparison of CSpeechSP and the Multi-Dimensional Voice Program Software Using the CSL 4300B Module and Multi-Speech for Windows☆

Stimuli 

The stimuli used in this study consisted of computer-generated fan noise mixed with voice samples. Fan noise was collected from a clinical laboratory computer (custom made 486 DX266 system) and recorded on a Tascam cassette recorder (112 MK II (TEAC Corporation, Montebello, CA)) using a Shure microphone (SM48 (Shure Brothers Incorporated, Evanston, IL)) that was positioned approximately 15 centimeters from the fan. The center frequency of the noise was approximately 235 Hz.

Subjects 

Voice samples from 10 female subjects, ages 19 to 25 (M=22) years old were used in this experiment. The average fundamental frequency of women's voices approached the center frequency of the noise. All subjects met the following criteria: (1) a history of normal speech and hearing (verified on the day of the recording), (2) normal health history as well as good health at the time of the experiment, (3) no history of smoking, (4) no professional voice training, (5) ability to match the fundamental frequency (F0) of a generated 235 Hz triangular tone (Hewlett Packard 33120A function/arbitrary waveform) within ±5 Hz, and (6) ability to maintain an acoustic intensity at 85 dBA ±2 dB (Bruel and Kjaer 2230 Sound Level Meter). Subjects passed a hearing screening before the voice recording took place (20 dB HL for the frequencies of 500, 1000, 2000, and 4000 Hz).

Procedure 

Voice acoustic signals were recorded in a sound-treated booth (Industrial Acoustics Corporation, Santa Cruz, CA). The noise stimulus was captured in a 20′×12′×9′ laboratory room. Ambient noise level in the booth was 35 dBA, and 58 dBA in the laboratory room (Bruel and Kjaer 2230 sound level meter).

The triangular wave (Hewlett Packard 33120A) used for pitch matching was amplified (Yamaha P2075) and delivered monaurally into the right earphone of a Telephonics TDH-50 headphone (after Gelfand19). Instrument calibration checks were performed daily (Quest Electronics 08-45 Octave Band Filter with a 9A coupler and a FC-3 frequency counter).

For the voice recording, a microphone (Shure SM 48) was positioned 10 cm from the mouth of each subject. The microphone of the sound level meter was positioned directly above the Shure microphone, and at the same distance to the mouth of the subjects. Each subject was instructed to sustain voiced /a/ for 10 seconds while matching the pitch, within ±5 Hz, of the generated triangular tone at a loudness level of 85 dBA. Each subject was also given the opportunity to monitor pitch by viewing the digital display of the frequency counter (Quest Electronics FC-3frequency counter). Ten productions per subject were recorded to a digital audiotape (DAT) recorder (Panasonic SV-255; 16-bit resolution at a sample rate of 48 kHz) located outside the booth.

Once recorded, each of the subject's 10 productions were analyzed directly from the DAT recorder through MDVP for Multi-Speech to verify that the subject met the F0 requirements, and that their relative average perturbation (RAP) and amplitude perturbation quotient (APQ) values were within “normal range” (RAP<0.68 and APQ<3.07, Kay Elemetrics, Lincoln Park, NJ).16 Peak-amplitude variation (vAm) was also considered in choosing samples. A single production from each subject was then selected that varied less than 2 Hz from an average of 235 Hz while maintaining the lowest RAP, APQ, and vAm values.

To establish signal-to-noise (S/N) ratio conditions, each selected voice sample and recorded noise sample were amplified (Yamaha P2075) and played through JBL speakers (Model Control 5). A microphone (Electro-Voice Model RE20) was placed 10 cm from the speaker source. The Electro-Voice microphone was used in place of the original recording microphone (Shure SM 48) to optimize impedance match between the microphone and the Creative Labs AWE64 sound card used to digitize signals from MDVP for Multi-Speech and CSpeechSP. The sound level meter was directed toward the speakers and positioned at approximately the same level as the microphone. Three different S/N levels were established (25, 20, and 15 dB S/N) and digitized directly into three systems. The S/N levels used were selected on the basis of clinical and laboratory experience with noise floors. The sound level of voice stimuli was verified (Bruel and Kjaer 2230 sound level meter) and maintained at 85 dB SPL, while noise levels were selected to create the S/N levels. Each voice stimulus was captured initially without the presence of the experimental noise. The nominal S/N level for this capture was 50 dB SPL or the amplitude of the voice signal (85 dB SPL) minus the noise floor in the booth (35 dB SPL). Mixed signals were then captured directly and sequentially by the following three computer arrangements, (1) MDVP for CSL 4300B (48 kHz sample rate), (2) MDVP for Multi-Speech (48 kHz sample rate), and (3) CSpeechSP (44.1 KHz sample rate). All samples had 16-bit resolution. In the CSpeechSP configuration, it was necessary to amplify the microphone signals using a Mic-Line Driver (FP-MLD). All three systems operated on the same computer (i.e., same mother board, same power supply, same monitor and monitor distance to sound boards). MDVP for Multi-Speech and CSpeechSP shared a Creative Labs Sound Blaster AWE64 audio card, whereas, MDVP for 4300B used the external Kay Elemetrics module for signal acquisition. The mixed signals were reduced to a 3-second mid-portion for subsequent processing.

Analysis 

Measures of F0, jitter, and shimmer were used in this study. Although CSpeechSP and the Kay Elemetrics software programs use different methods for calculating these three measures, the acoustic signals were analyzed only using CSpeechSP. Thus, all measures were derived from the least mean squares formula.11 Signals captured into the Kay Elemetrics software and hardware arrangements (MDVP for CSL 4300B and for Windows) were imported to CSpeechSP. A comparison between the waveforms of the original and imported versions showed no file corruption. The CSpeechSP program expresses jitter in milliseconds and shimmer in percent.

Repeated Measures MANOVAs were used to determine significant within subject differences on each variable (F0, jitter, and shimmer) by software programs/hardware arrangements (device) and by S/N levels (condition) (SPSS 9.0 for Windows). Within subjects post hoc comparisons with associated Bonferroni20 adjustments were used when significant main effects were found. Dependent t-tests with associated Bonferroni adjustments were used when significant interaction effects were noted. An alpha level of <0.05 was established.

Fundamental frequency 

The multivariate test for device was not significant (Wilks' lambda=0.54, F=3.43, p=0.08). This finding indicated that F0 values were comparable across the three software programs (see Table 1). The test for S/N level was significant (Wilks' lambda=.27, F=6.42, P=0.02). Post hoc comparisons revealed that F0 recorded with no additive environmental noise was significantly higher than the worst S/N condition, 15 dB.

Jitter 

Significant main effects for device (Wilks' lambda=0.42, F=5.43, P=0.03) and condition (Wilks' Lambda=0.07, F=32.32, P=<0.001) were found; however, there was a significant interaction effect (Wilks' lambda=0.09, F=6.66, P=0.04). Results from dependent t-tests comparing jitter values at different S/N levels per computer software system are displayed in Table 2. Those results revealed that: (a) For the MDVP 4300B unit, significant differences were noted between jitter values recorded with no additive noise (50 dB) and those at 15 dB S/N, between 25 dB and 20 dB S/N, between 25 dB and 15 dB S/N, and between 20 dB and 15 dB S/N; (b) For MDVP operating with Multi-Speech three of the jitter comparisons were significant; those differences were noted between jitter values recorded with no additive environmental noise (50 dB) and those at 15 dB S/N, between 25 dB and 15 dB S/N, and between 20 dB and 15 dB S/N; (c) for the CSpeechSP program, four of the six comparisons were significant: those recorded with no additive noise (50 dB) versus 20 dB S/N, no additive noise (50 dB) versus 15 dB S/N, 25 versus 20 dB S/N, and 25 versus 15 dB S/N. In all cases, jitter values recorded with additive noise were higher than those estimated without the presence of additive environmental noise. As shown in Table 3, only one significant difference was found when making paired comparisons of the three computer-ized voice analysis systems at each of the S/N condi-tions. Jitter values recorded on MDVP for CSL 4300B were significantly higher than those recorded on CSpeechSP at 15 dB S/N (see Table 1 for mean values).

TABLE 2. Dependent t-Tests Comparing Jitter Values Across S/N Conditions with Significance Level Adjusted by the Bonferroni Procedure

	Device	S/Na pair	t-value (dfb=9)	Significance
	MDVP for CSL 4300B	50 dBc vs. 25 dB	−2.76	0.396
		50 dB vs. 20 dB	−3.65	0.090
		50 dB vs. 15 dB	−4.36	0.036∗
		25 dB vs. 20 dB	−4.49	0.036∗
		25 dB vs. 15 dB	−4.80	0.018∗
		20 dB vs. 15 dB	−4.74	0.018∗
	MDVP for Multi-Speech	50 dB vs. 25 dB	−3.00	0.270
		50 dB vs. 20 dB	−4.13	0.054
		50 dB vs. 15 dB	−5.02	0.018∗
		25 dB vs. 20 dB	−3.79	0.072
		25 dB vs. 15 dB	−5.08	0.018∗
		20 dB vs. 15 dB	−5.25	0.018∗
	CSpeechSP	50 dB vs. 25 dB	−2.86	0.342
		50 dB vs. 20 dB	−4.36	0.036∗
		50 dB vs. 15 dB	−4.40	0.036∗
		25 dB vs. 20 dB	−6.00	>0.001∗
		25 dB vs. 15 dB	−4.50	0.018∗
		20 dB vs. 15 dB	−4.12	0.054

Note: 25 dB is the best S/N condition created, while 15 dB is the worst S/N condition.

a S/N=Signal-to-noise ratio measured in decibels A-scale. b df=Degrees of freedom. c 50 dB=No additive noise was added to this signal. ∗ p<0.05.

TABLE 3. Dependent t-Tests Comparing Jitter Values Across Deviceswith Significance Level Adjusted by the Bonferroni Procedure

	S/Na Condition	Devices Pair	t-value (dfb=9)	Significance
	50 dBc	4300B vs. Multi-Speech	−2.45	0.444
		4300B vs. CSpeechSP	−0.32	9.096
		Multi-Speech vs. CSpeechSP	1.41	2.316
	25 dB	4300B vs. Multi-Speech	1.50	2.016
		4300B vs. CSpeechSP	1.77	1.332
		Multi-Speech vs. CSpeechSP	−0.56	7.092
	20 dB	4300B vs. Multi-Speech	2.11	0.768
		4300B vs. CSpeechSP	2.43	0.456
		Multi-Speech vs. CSpeechSP	1.50	2.016
	15 dB	4300B vs. Multi-Speech	3.20	1.332
		4300B vs. CSpeechSP	4.05	0.036∗
		Multi-Speech vs. CSpeechSP	2.57	0.350

Note: Mixed signals were captured separately on the Multi-Dimensional Voice Program (MDVP) software option for CSL 4300B, MDVP for Multi-Speech, and CSpeechSP. To obtain jitter values, signals originally captured on MDVP for CSL 4300B and MDVP for Multi-Speech were imported and analyzed on CSpeechSP.

a S/N=Signal-to-noise ratio measured in decibels, A-scale. b df=Degrees of freedom. c 50 dB=No additive noise was added to this signal. ∗ p<0.05.

Shimmer 

Main effects and the interaction effect were significant for shimmer values (device, Wilks' lambda=0.11, F=32.55, P=>0.001; condition, Wilks' lambda=0.01, F=227.55, P=>0.001; device by condition, Wilks' lambda=0.08, F=7.39, P=0.04). Post hoc testing, comparing shimmer values recorded per device at each of the S/N conditions, revealed that all comparisons were significantly different (see Table 4). Significant differences were found on all comparisons between the varying S/N conditions, regardless of the computerized voice analysis system. The trend was as noise floors increased, shimmer values increased, regardless of the analysis system.

TABLE 4. Dependent t-Tests Comparing Shimmer Values Across S/N Conditions with Significance Level Adjusted by the Bonferroni Procedure

	Device	S/Na pair	t-value (dfb=9)	Significance
	MDVP for CSL 4300B	50 dBc vs. 25 dB	−20.19	<0.001∗
		50 dB vs. 20 dB	−20.59	<0.001∗
		50 dB vs. 15 dB	−25.70	<0.001∗
		25 dB vs. 20 dB	−13.81	<0.001∗
		25 dB vs. 15 dB	−25.42	<0.001∗
		20 dB vs. 15 dB	−21.50	<0.001∗
	MDVP for Multi-Speech	50 dB vs. 25 dB	−13.73	<0.001∗
		50 dB vs. 20 dB	−16.85	<0.001∗
		50 dB vs. 15 dB	−29.10	<0.001∗
		25 dB vs. 20 dB	−14.05	<0.001∗
		25 dB vs. 15 dB	−30.19	<0.001∗
		20 dB vs. 15 dB	−29.56	<0.001∗
	CSpeechSP	50 dB vs. 25 dB	−11.66	<0.001∗
		50 dB vs. 20 dB	−14.31	<0.001∗
		50 dB vs. 15 dB	−18.18	<0.001∗
		25 dB vs. 20 dB	−14.69	<0.001∗
		25 dB vs. 15 dB	−19.64	<0.001∗
		20 dB vs. 15 dB	−19.33	<0.001∗

Note: 25 dB is the best S/N Condition created (containing the least amount of noise), while 15 dB is the worst S/N Condition (containing the most amount of noise).

a S/N=Signal-to-noise ratio measured in decibels, A-scale. b df=Degrees of freedom. c 50 dB=No additive noise was added to this signal. ∗ p<0.05.

When making pair-wise comparisons among the devices with S/N level held constant, as shown in Table 5, significant differences were noted between the MDVP for the 4300B unit and MDVP for Multi-Speech at 25 dB and 15 dB S/N conditions. In both instances, shimmer values recorded from the 4300B unit were higher than those from MDVP for Multi-Speech, which can be seen in Table 1. Note that differences between these two devices at the 20 dB S/N level approached significance (P=0.06). Likewise, values taken from voice analyses using MDVP for the 4300B unit were significantly higher than those recorded from CSpeechSP at the 25, 20, and 15 dB S/N levels.

TABLE 5. Dependent t-Tests Comparing Shimmer Values Across Devices with Significance Level Adjusted by the Bonferroni Procedure

	S/Na Condition	Devices Pair	t-value (dfb=9)	Significance
	50 dBc	4300B vs. Multi-Speech	−1.01	4.080
		4300B vs. CSpeechSP	−0.09	11.172
		Multi-Speech vs. CSpeechSP	2.54	0.384
	25 dB	4300B vs. Multi-Speech	4.63	0.012∗
		4300B vs. CSpeechSP	6.26	<0.001∗
		Multi-Speech vs. CSpeechSP	1.26	2.868
	20 dB	4300B vs. Multi-Speech	3.73	0.050
		4300B vs. CSpeechSP	7.76	<0.001∗
		Multi-Speech vs. CSpeechSP	1.41	2.292
	15 dB	4300B vs. Multi-Speech	5.48	<0.001∗
		4300B vs. CSpeechSP	8.75	<0.001∗
		Multi-Speech vs. CSpeechSP	2.97	0.192

Note: To compare devices, mixed signals were captured separately on the Multi-Dimensional Voice Program (MDVP) software option for CSL 4300B, MDVP for Multi-Speech, and CSpeechSP. To obtain shimmer values, signals originally captured on MDVP for CSL 4300B and MDVP for Multi-Speech were imported and analyzed on CSpeechSP.

a S/N=Signal-to-noise ratio measured in decibels, A-scale. b df=Degrees of freedom. c 50 dB=No additive noise was added to this signal. ∗ P<0.05.