Acoustical Comparison Between Samples of Good

and Poor Vibrato in Singers

*Jose A. Diaz and †Howard B. Rothman

*Carabobo, Venezuela
†Gainesville, Florida

Summary: The purpose of this research was to analyze samples of frequency

vibrato taken from recordings of eight different singers, which were classified
as examples of good or poor singing. The samples were analyzed by a software
package, which makes use of the linear prediction coding (LPC) method to
determine the time varying rate and extent of the frequency vibrato wave.
Four parameters, which relate to the periodicity of the samples, were extracted
from the time varying rate and extent and investigated in order to verify or
reject the hypothesis that the best vibrato samples were the most symmetric
ones. Ten samples per singer were analyzed, 5 good and 5 poor, for a total
of 80 samples. The results show that the samples judged as good were the
most periodic ones.
Key Words: Vibrato—Singers’ voice—Speech Processing—Spectrogram—
Linear Prediction Coding (LPC).

INTRODUCTION pedagogues, vocal critics, and scientists who study

Historically, in order to determine the quality of
a vibrato sample, experts in the area of vocal music
and the singing voice would listen to, analyze, and
judge the quality of vibrato samples.1 This is a
subjective process; however, there is agreement
among experts. Most experts, that is, singers, vocal
Accepted for publication August 9, 2002.
Presented at the II Iberoamerican Congress on Acoustics held
in Madrid from October 16 to 20, 2000.
From the *Department of Electrical Engineering, University
of Carabobo and †Department of Communication Sciences and
Disorders, University of Florida, Gainesville, Florida.
Address correspondence and reprint requests to Jose A. Diaz,
Department of Electrical Engineering, University of Carabobo,
Venezuela, P.O. Box 025685, Miami, FL 33102-5685. E-mail:
jadiaz@thor.uc.edu.ve
Journal of Voice, Vol. 17, No. 2, pp. 179–184
쑕 2003 The Voice Foundation
0892-1997/2003 $30.00⫹0
doi:10.1016/S0892-1997(03)00002-X
vibrato, agree that the samples with better quality
are the most periodic ones,2 that is, the ones whose
rate and extent are more constant and regular
over time.
The main objective of this research was to study
in detail the periodicity features of vibrato samples,
in order to verify or reject the hypothesis that the
best samples were the most periodic ones. In order
to accomplish this objective, signal processing al-
gorithms and software were used to extract and
analyze some parameters related to the periodicity
of the vibrato wave. There was also an expectation
that the parameters directly related to vibrato quality
could be identified. Therefore, this research presents
a novel approach to vibrato analysis because there
does not appear to be any other analysis being per-
formed to measure vibrato in singers or musical
instruments in which the time varying rate and extent
of the vibrato samples were analyzed at a resolution
as good or better than 80.58 samples per second, nor
through the use of linear prediction coding (LPC).3–9

179
180 JOSE A. DIAZ AND HOWARD B. ROTHMAN
FIGURE 1. Spectrogram of vibrato sample.
SAMPLE SELECTION
For this research, eight professional singers (5
male, 3 female) with years of experience were
chosen from a database of vibrato samples that is kept
at the Department of Communication Sciences and
Disorders at the University of Florida. Four different
individuals with singing experience or with experi-
ence in studying the singing voice listened to all
samples from the eight singers in order to classify
the samples produced by each singer as samples of
good or poor vibrato. The judges were instructed to
classify the samples depending on their quality, as
good (pleasing to the ear) or poor (disagreeable to
the ear).
The results provided by each judge were com-
pared, and the samples that were judged as good
or poor by the four judges were chosen for analysis;
the other samples were rejected because the
judges did not unanimously agree on their quality.
From the chosen samples, five samples of good vi-
brato and five of poor vibrato were randomly se-
lected for each singer, which resulted in a total of
80 samples (8 singers times 10 samples/singer). All
samples selected fulfilled the following criteria:
1. Length greater than 1.5 s.
2. No change in pitch.
3. No change in vowel.
4. Absence of accompanying musical instru-
ments or noise.
SAMPLE ANALYSIS
The MMSV1 (Mathematical Model of Singers’
Vibrato, The Mathworks, Natick, MA) software,
Journal of Voice, Vol. 17, No. 2, 2003
FIGURE 2. Example of frequency vibrato.
which was developed by one of the authors, was
selected for sample analysis because it provided
the time varying rate and extent in Hertz of the
frequency vibrato wave. This software generated
sample spectrograms (see Figure 1), from which the
vibrato samples were extracted. Figure 2 shows
the frequency vibrato wave extracted from the sixth
harmonic in Figure 1.
The software used the LPC method to generate the
time varying rate and extent waves of the frequency
vibrato, which represent how the rate and extent of
the frequency vibrato wave vary along the time axis.
These two waves were created by taking a small
segment of the vibrato signal (170 ms.), applying
the LPC method, and obtaining the rate and extent
in Hertz of the segment from the LPC parameters
obtained. Then, the software selected a new segment
of the same length (170 ms.), shifted 12 ms. to the
right and analyzed it. This process was repeated
until the entire signal was analyzed.
The length of the vibrato segment was chosen to
be equal to 170 ms. because a longer segment would
show the average rate and extent of the segment,
and a shorter segment would not contain enough
data for a reliable result. Then, all rate and extent
values were plotted on separate windows (see Fig-
ures 3 and 4). These two waves represent the time
varying rate and extent of the frequency vibrato
wave seen in Figure 2. Notice that the time scale
in Figures 3 and 4 is shorter than that of Figures 1
and 2. This is because the LPC method requires a
segment of the vibrato wave to calculate one value
of the rate and extent. It is important to note that
the time varying extent wave represents the extent
of the wave in Figure 2, and it does not represent
 ACOUSTICAL COMPARISON OF GOOD AND POOR VIBRATO 181

the amplitude vibrato wave, which contains the am-

plitude variations associated with the frequency
vibrato.
The main advantage in using this software com-
pared to other methods was the time resolution
(number of samples per second), which is 80.58
samples per second in Figures 3 and 4. This allowed
the variation patterns in rate and extent of the fre-
quency vibrato to be observed and compared.
If the frequency vibrato wave was a sinusoidal
wave, the waves in Figures 3 and 4 would be straight FIGURE 3. Time varying rate of frequency vibrato.
lines because the rate and extent of a pure sine wave
are constant over time. Therefore, as the waves in
Figures 3 and 4 approach a straight line, the fre-
quency vibrato becomes more periodic, and the
departure from a straight line shape indicates the
degree of aperiodicity in the frequency vibrato
wave.
As a result, in order to measure the periodicity
of a vibrato sample, four parameters were selected
that measure the deviations of the waves in Figures
3 and 4 from a straight line. These parameters are
as follows:
FIGURE 4. Time varying extent of frequency vibrato.
1. Total time varying rate deviation (maximum
rate minus minimum rate) in Hertz.
2. Variability (standard deviation) of the time
varying rate in Hertz.
3. Total time varying extent deviation (maximum The 80 samples chosen for analysis were analyzed
extent minus minimum extent) in Hertz. using the MMSV software in order to generate the
4. Variability (standard deviation) of the time time varying rate and extent waves for each of them.
varying extent in Hertz. Then, the four parameters mentioned above were
Parameters 1 and 2 were extracted from the time calculated by a Matlab-based program (The Math-
varying rate wave and parameters 3 and 4 were ex- works, Natick, MA), which also saved the parame-
tracted from the time varying extent wave. Param- ters in an Excel-compatible format for later analysis.
eters 2 and 4 (variability) represent the spread of
the time varying rate and extent in Hertz.
In order to make a meaningful comparison among
samples and singers, the four parameters mentioned RESULTS
above were converted into percentages of the aver- Tables 1 through 8 show the results of the four
age value of the wave from which the parameter was parameters under analysis for the eight singers after
extracted. This was necessary because the MMSV they were converted into percentages as explained
software can extract the vibrato wave from different above. The first column indicates the parameter
harmonics; therefore, for higher harmonics, the name. Columns 2 and 3 show the average values ob-
variability is higher and it should not be compared tained for the good and poor samples, respectively.
to a time varying wave extracted from a lower Single-factor analyses of variance10 (ANOVAs) with
harmonic. By converting each parameter into five samples per group, and two groups (good and
percent of the average wave value, this error was poor samples) were applied to the four parameters
canceled. of the eight singers (critical value of F = 5.32

Journal of Voice, Vol. 17, No. 2, 2003

182 JOSE A. DIAZ AND HOWARD B. ROTHMAN

TABLE 1. Results for Singer Number 1

Avg good samp Avg poor samp F value p value Significant

Time varying rate variability % 4.1672 5.388 4.403696 0.069099 No

Time varying rate total dev % 19.48266 25.12942 3.369403 0.103744 No
Time varying ext variability % 6.7132 7.9686 0.74747 0.41245 No
Time varying ext total dev % 49.85025 56.05383 0.341805 0.574902 No

TABLE 2. Results for Singer Number 2

Avg good samp Avg poor samp F value p value Significant

Time varying rate variability % 5.2718 5.438 0.046982 0.833826 No

Time varying rate total dev % 23.07027 26.39093 0.652133 0.442698 No
Time varying ext variability % 5.4446 13.8324 13.61278 0.006134 Yes
Time varying ext total dev % 38.4382 100.0234 16.93521 0.003366 Yes

TABLE 3. Results for Singer Number 3

Avg good samp Avg poor samp F value p value Significant

Time varying rate variability % 6.6054 10.4638 2.911387 0.126346 No

Time varying rate total dev % 30.66759 48.50149 3.11204 0.115722 No
Time varying ext variability % 7.555 13.9394 8.675875 0.018553 Yes
Time varying ext total dev % 56.5322 103.4366 9.830369 0.013903 Yes

TABLE 4. Results for Singer Number 4

Avg good samp Avg poor samp F value p value Significant

Time varying rate variability % 4.5234 5.2498 0.372504 0.558578 No

Time varying rate total dev % 21.00638 23.78328 0.22438 0.64838 No
Time varying ext variability % 9.1682 11.8154 2.435761 0.157218 No
Time varying ext total dev % 68.10582 89.01863 2.424016 0.158102 No

TABLE 5. Results for Singer Number 5

Avg good samp Avg poor samp F value p value Significant

Time varying rate variability % 3.3768 4.3642 3.630769 0.093181 No

Time varying rate total dev % 15.50761 21.98027 3.753822 0.088691 No
Time varying ext variability % 4.5994 6.3954 7.808637 0.023399 Yes
Time varying ext total dev % 37.0789 47.01511 2.875563 0.128376 No

for α = 0.05, df1 = 1, and df2 = 8). Columns 4 and
5 show the F and P values respectively. No arc-
sine transform was performed on the data. Column 6
indicates whether the obtained results were significant.
Significant results were found in four of the eight
singers, which indicates that the mean values of
the variables under study were higher for the poor
samples in four singers of eight. No significant re-
sults were found for singers numbered 1, 4, 6, and
8. These results support the hypothesis that the best
samples were the most symmetrical ones and serve
to validate the listeners’ classification of the samples.

Journal of Voice, Vol. 17, No. 2, 2003

 ACOUSTICAL COMPARISON OF GOOD AND POOR VIBRATO 183

TABLE 6. Results for Singer Number 6

Avg good samp Avg poor samp F value p value Significant

Time varying rate variability % 3.1004 4.3288 1.997513 0.195265 No

Time varying rate total dev % 14.52671 19.16041 1.05888 0.333575 No
Time varying ext variability % 6.2266 10.4394 5.171822 0.052548 No
Time varying ext total dev % 48.69182 71.93773 3.480643 0.099067 No

TABLE 7. Results for Singer Number 7

Avg good samp Avg poor samp F value p value Significant

Time varying rate variability % 7.1548 7.471 0.032922 0.860532 No

Time varying rate total dev % 33.52693 32.18296 0.029144 0.868687 No
Time varying ext variability % 6.5218 10.9056 6.52134 0.033979 Yes
Time varying ext total dev % 46.75865 78.79748 6.500981 0.034191 Yes

TABLE 8. Results for Singer Number 8

Avg good samp Avg poor samp F value p value Significant

Time varying rate variability % 4.1346 4.6404 0.148622 0.709904 No

Time varying rate total dev % 17.21066 29.39715 1.006298 0.345165 No
Time varying ext variability % 6.4406 7.2862 0.617938 0.454466 No
Time varying ext total dev % 44.56048 55.2096 1.728126 0.225085 No

Analysis of the vibrato rate for the
individual singers
We can see from Tables 1 through 8 that the
variability in rate for the poor samples was always
higher than that of the good samples. The total devia-
tion of the time varying rate was higher in seven
of the eight singers and was only smaller for singer
7. The higher total deviation of the time varying
rate for the good samples in singer 7 indicates large
oscillations of the time varying rate, which were not
perceived as poor vibrato. The statistical analysis
of the variability of the time varying rate did not
show significant results, nor did the total deviation
of the time varying rate.
Analysis of the vibrato extent for the
individual singers
We can see from Tables 1 through 8 that the
variability and the total deviation of the time varying
extent for the poor samples were always higher than
that of the good samples. The statistical analysis
of the variability of the time varying extent yielded
significant results in four singers of eight, whereas
the total deviation of the time varying extent yielded
significant results in three cases of eight.
Comparison between good and poor samples
for the entire group
Table 9 provides a summary of the results after
comparing the good and poor samples of the entire
group. The first column shows the parameters
under analysis. Columns 2 and 3 show the average
results for good and poor samples. Single factor-
ANOVAs with eight samples per group (8 singers)
and two groups (good and poor samples) were ap-
plied to the four parameters under analysis (critical
value of F = 4.60 for α = 0.05, df1 = 1, and df2 =
14). Columns 4 and 5 show the F and p values
respectively. No arc-sine transform was performed
on the data. Column 6 indicates whether the ob-
tained results were significant or not.

Journal of Voice, Vol. 17, No. 2, 2003

184 JOSE A. DIAZ AND HOWARD B. ROTHMAN
TABLE 9. Comparison Between Good and Poor Samples
Avg good samp
Time varying rate variability % 4.7918
Time varying rate total dev % 21.87485
Time varying ext variability % 6.583675
Time varying ext total dev % 48.75204
Analysis of the vibrato rate for the
entire group
Table 9 indicates that the mean values of the
four parameters were always smaller for the good
samples than for those of the poor samples. In Table
9, the variability and the total deviation of the time
varying rate do not show significant results. How-
ever, it is important to note that the poor samples
used in this study showed higher oscillations in rate
than that of the good samples, although no significant
results were found.
Analysis of the vibrato extent for the
entire group
In Table 9, the variability and the total deviation
of the time varying extent show significant results.
Therefore, it appeared that the oscillations observed
in the rate and extent of the frequency vibrato wave
were mainly due to variations in extent.
CONCLUSIONS
The results of this research support the hypothesis
that the most symmetrical samples were judged
as good samples of vibrato. Also, it appeared that
the largest number of significant results was found
in the variability of the time varying extent (4 cases
of 8), which suggests that the variability of the time
varying extent provided a measure of vibrato quality
that was more accurate and meaningful than the
other three parameters under analysis.
Journal of Voice, Vol. 17, No. 2, 2003
Avg poor samp F value p value Significant
5.918 1.562169 0.231838 No
28.31574 2.527292 0.134214 No
10.3228 10.99749 0.005095 Yes
75.18655 10.06189 0.006788 Yes
It could also be seen that the oscillations in the
rate and extent of frequency vibrato wave were
found mainly in the time varying extent wave, rather
than in the time varying rate wave, which shows that
it was more difficult for singers to control the extent of
the vibrato pulse than the rate at which the vibrato
pulse occurred. Taken together, these results show
a direct relationship between the periodicity of the
vibrato wave and its perceived quality.
REFERENCES
1. Diaz JA. Frequency characterization of singers’ vibrato.
Gainesville, FL: University of Florida; 1995. Master’s thesis.
2. Sundberg J. The Science of the Singing Voice. Dekalb, IL:
Northern Illinois University Press; 1987.
3. DeJonckere PH, Minoru H, Sundberg J. Vibrato. San Diego,
California: Singular Publishing Group; 1995.
4. Prame E. Measurement of the vibrato rate of ten singers.
J Acoust Soc Am. 1994;96:1979–1984.
5. Horii Y. Frequency modulation characteristics of sustained
/a/ sung in vocal vibrato. J Speech Hear Res. 1989;32:1–8.
6. Maher R, Beauchamp J. An investigation of vocal vibrato
for synthesis. Appl Acoustics. 1990;30:19–245.
7. Sundberg J. Acoustic and psychoacoustic aspects of vocal
vibrato. KTH Speech Transmission Lab Quart Progr Status
Report. 1994;2–3:45–48.
8. Titze IR. Synthesis of sung vowels using a time-domain
approach. Transcripts of the Eleventh Symposium: Care of
the Professional Voice. New York: The Voice Foundation;
1983:90–98.
9. Mellody M, Wakefield G. The time-frequency character-
istics of violin vibrato: modal distribution analysis and
synthesis. J Acoust Soc Am. 2000;107:598–611.
10. Ott RL. An Introduction to Statistical Methods and Data
Analysis. Belmont, CA: Duxbury Press; 1993.