Significance of analysis window size in maximum flow declination rate (MFDR)

Linda M. Carroll, PhD
Department of Otolaryngology, Mount Sinai School of Medicine

Goal:

1. To determine whether a significant difference exists for mean MFDR across 4 different data extraction methods on the same data set.
2. To determine interaction between subject skill level and fundamental frequency on MFDR.

Background:

Examination of laryngeal aerodynamics remains crucial to our understanding of voice function in normal and non-normal subjects. Extensive research over the past 40 years has focused on subglottal pressure and transglottal flow, particularly as it relates to frequency and intensity control. More recently, the speed of closure at the maximal negative slope of the differentiated inverse-filtered waveform, or maximum flow declination rate (MFDR), has emerged as a valuable measure of laryngeal function (1-8). Although subglottal pressure and transglottal flow have established measurement techniques for data extraction methods (e.g.: peak pressure value during [p] for subglottal pressure), such standards do not exist for MFDR. As such, it becomes difficult to compare results across studies which have used a wide range of measurement techniques.

Assumptions

• MFDR is the point of sharpest change in the closing velocity of the vocal folds, and reflects the velocity when the vocal fold surfaces are nearly parallel and touching in the anterior (membranous) glottis (1-4, 9).
• It is hypothesized that a more rapid decrease (or stoppage) of the flow yields a more efficient and powerful glottal source, thereby allowing improved acoustic intensity (3,4,5,7,9).
• Previous investigators have reported MFDR values for speaking and singing using a range of 1-60 periods of analysis (3,6,7,10,11).

Experimental Design:

Subjects:

• Eight professional lyric sopranos employed as solo artists at international opera houses (N=4) or regional/national opera houses (N=4) served as volunteers in the IRB-approved study.

Tasks:

• Three tokens of a 7-syllable /pa/ train at progressively increasing and then decreasing intensity (messa di voce) in singing mode at two contrasting frequencies (F01=330 Hz, F02=660 Hz), with each /pa/ syllable lasting 1 second in duration.
• Intensity changes were not prescribed. The subject was instructed to sing a messa di voce as they typically would on the operatic stage.

Figure 1. Sample flow waveform for subject during 7-syllable /pa/ task.

Data Collection:

The subject held a pneumotachograph mask firmly in place over her nose and mouth, with a pressure tube passing between the lips. A microphone was fitted in the mask handle.

• Signals from flow, pressure and microphone were digitized by a 12-bit analog to digital converter board with a sampling rate of 10 kHz per channel. Digitized signals were imported to the Alamed Voice Plus? and CSpeech? 3.1 analysis systems on a Pentium based computer. Waveforms were optimized by adjusting the amplifier gain to ensure optimum signal input for each subject prior to data collection, and were monitored by a Tektronix? TDS-420A 4-channel digitizing oscilloscope during computer data collection.
• Glottal velocity waveforms were recorded from two differential pressure transducers (Glottal Enterprises? PTL-2) mounted in a Rothenberg single-layer circumferentially vented pneuomotachograph mask, which was connected to a Glottal Enterprises? MSIF-2 inverse-filtering unit.
• Calibration for pressure (water u-tube manometer) and airflow (Matheson glass-float rotameter) was done immediately after each subject’s data collection using known pressures and flow that produced output voltages that approximated those observed on the oscilloscope during data collection.

Figure 2. Block diagram of experimental instrumentation.

Data Analysis:

• The most negative value from the first derivative of the inverse-filtered waveform (MFDR) was extracted using CSpeech? 3.1 for each cycle at F01 and F02.
• Some subjects’ differentiated flow waveforms had two negative peaks, which were often reduced to one negative peak 20 ms later, as a result of unexpected, intermittent presence of formant energies from high voice quality. Because of errors in peak detection for automated MFDR computation, hand cycle-by-cycle determination of the MFDR value for each cycle was used for the eight subjects.
  • Raw flow signal was compared with the inverse-filtered flow signal during MFDR detection at F01 (330 Hz) and F02 (660 Hz) for most negative point within each cycle.

Figure 3. Sample flow signal (A) and inverse filter of signal (B).

Figure 4. Sample of Inverse-filtered flow signal with differentiated waveform for MFDR for one subject. Tracings show easy marking of MFDR point for upper trace, and need for hand-marking of MFDR point in lower trace with change in cursor position within /pa/ from 256 ms into /pa/ (upper trace) to 321.7 ms into /pa/ (lower trace).

Subject performance was compared from 4 different extraction windows within each /pa/ for the 7-syllable train at F01 and F02.

• Method A: mean MFDR from middle 1000 ms for each /pa/ segment (if less than 1000 ms available in /pa/, then 20 ms excluded from onset/offset)
  • analysis of 330 cycles for F01, 660 cycles for F02 at mid-portion
• Method B: mean MFDR from middle 100 ms of /pa/ segment, with center at mid-portion of entire /pa/ segment
  • analysis of 33 cycles for F01, 66 cycles for F02 at mid-portion
• Method C: mean MFDR for -/+ 50 ms from greatest value of MFDR from entire /pa/ segment
  • analysis of 33 cycles for F01, 66 cycles for F02 at greatest value
• Method D: mean MFDR for -/+ 10 cycles from greatest value of MFDR from entire /pa/ segment
  • 20 cycles for F01, 20 cycles for F02 at greatest value

Statistical Analysis

SPSS?®, with overall=0.05 , with each /pa/ studied as unique variables. Each subject’s mean MFDR (and sd-MFDR) was a composite of three trial tokens at each pitch condition.

• Analyses of variance (ANOVA) were used to test whether a significant difference exists for MFDR across the four different measurement techniques. Statistical adjustment was made for pitch, group, and all interactions. These analyses were repeated for each /pa/ during the 7-syllable train.
• In the ANOVAs, pitch, group and window were fixed factors, and subjects within groups was a random factor. We used a full ANOVA model that included all interactions. A significance level of 0.05 was used for each analysis. Marginal significance was defined as a p-value between 0.05 and 0.10. Contrasts were performed to compare the four measurement techniques for each /pa/. Bonferroni adjustments were used for these pairwise comparisons, with an overall =0.05.

Results:

• Mean MFDR
  • A significant main effect was found for pitch condition (F01, F02) at [pa3], [pa4] and [pa5]
  • A significant main effect was found for window (method 1, 2, 3, 4) at [pa1] through [pa6], with marginal main effect at [pa7]
  • A significant pitch condition by group (A, B) interaction was found at [pa1] through [pa6]
  • A significant pitch condition by window interaction was found at [pa4] and [pa5], with marginal 2-way interaction for [pa3]
• Standard deviation of MFDR (sd-MFDR)
  • A significant main effect was found for pitch condition throughout the 7-syllable /pa/ train
  • A significant main effect was found for window at [pa1] through [pa6]
  • A marginal main effect was found for group at [pa2]
  • A significant pitch condition by group interaction was found throughout the 7-syllable /pa/ train
  • A significant pitch condition by window interaction was found at [pa3] through [pa5], with marginal 2-way interaction at [pa2]
• No significant difference between groups for MFDR or sd-MFDR at any [pa]
• Pairwise comparisons
  • Window 1 vs. window 2
    • No significant difference between mean MFDR
    • Significantly different sd-MFDR for window 2 at [pa1], [pa2], [pa5] and [pa6]
  • Window 1 vs. window 3
    • Significantly greater mean MFDR for window 3 at [pa2], [pa5] and [pa6], with marginal significance at [pa1] and [pa3]
    • Marginally different sd-MFDR for window 3 at [pa5]
  • Window 2 vs. window 3
    • No significant difference for mean MFDR or sd-MFDR
  • Window 3 vs. window 4
    • No significant difference for mean MFDR
    • Significant difference for sd-MFDR at [pa2], [pa3], [pa5], and [pa6], with marginal significance at [pa4]

Table 1. MFDR means (and standard deviations) and maximum MFDR for F01 (pitch=1) for four different data extractions

	Group	Peak Value	M1 (in l/s)	M2 (in l/s)	M3 (in l/s)	M4 (in l/s)
Pa1	A B	488 264	79 (25.71) 75 (29.64)	83 (10.33) 83 (11.23)	105 (18.51) 113 (16.72)	120 (13.67) 118 (15.91)
Pa2	A B	547 166	122 (30.56) 67 (19.16)	129 (13.31) 74 (7.17)	155 (14.46) 91 (9.81)	158 (14.1) 93 (9.06)
Pa3	A B	313 283	145 (22.97) 97 (23.63)	150 (17.76) 108 (10.18)	154 (19.36) 125 (11.26)	173 (16.44) 128 (12.56)
Pa4	A B	430 264	157 (21.48) 114 (22.14)	159 (17.52) 121 (10.66)	175 (19.88) 133 (11.4)	181 (21.07) 135 (11.44)
Pa5	A B	254 596	124 (18.97) 109 (30.23)	132 (11.24) 129 (12.58)	143 (13.49) 147 (17.22)	146 (12.29) 151 (16.54)
Pa6	A B	195 127	67 (11.49) 52 (19.23)	74 (6.75) 54 (7.8)	79 (7.22) 76 (7.98)	81 (6.78) 79 (7.49)
Pa7	A B	88 195	36 (8.2) 27 (10.61)	37 (4.75) 31 (6.68)	43 (5.22) 41 (7.9)	44 (4.8) 43 (7.78)

Table 2. MFDR means (and standard deviations) and maximum MFDR for F02 (pitch=2) for four different data extractions

	Group	Peak Value	M1 (in l/s)	M2 (in l/s)	M3 (in l/s)	M4 (in l/s)
Pa1	A B	1436 264	167 (114.74) 48 (49.38)	210 (94.82) 50 (20.03)	315 (121.77) 85 (32.8)	422 (50.86) 116 (9.79)
Pa2	A B	908 352	178 (120.43) 44 (30.82)	242 (91.88) 50 (23.53)	310 (115.32) 85 (36.67)	396 (50.39) 122 (12.84)
Pa3	A B	1221 986	233 (116.1) 91 (65.81)	250 (95.3) 105 (46.47)	345 (109.87) 170 (77.86)	442 (52.73) 240 (39.13)
Pa4	A B	1084 811	238 (130.74) 112 (72.39)	264 (126.2) 154 (57.64)	320 (124.77) 199 (75.34)	445 (51.48) 274 (38.93)
Pa5	A B	1191 908	201 (119.57) 87 (50.12)	266 (111.25) 108 (39.65)	344 (92.44) 152 (52.89)	415 (42.52) 204 (28.76)
Pa6	A B	430 352	78 (40.63) 47 (30.86)	92 (40.41) 54 (18.67)	118 (45.71) 79 (28.94)	160 (13.49) 109 (11.51)
Pa7	A B	264 186	45 (23.65) 28 (13.69)	47 (20.97) 35 (11.32)	66 (25.98) 43 (13.05)	89 (11.05) 56 (7.37)

Table 3. Main effect and interaction effects of pitch condition (frequency F01, F02), window size (method 1, 2, 3, 4) and group (A, B) on mean MFDR and sd-
MFDR for each individual /pa/ during the 7-syllable train.*

	Source	F	Sig.=0.05		Source	F	Sig.=0.05
Pa1	Pitch Window Group Pitch x window Pitch x group Window x group	0.507 5.212 0.782 0.638 22.006 0.005	ns 0.004** ns ns 0.000*** ns	sdPa1	Pitch Window Group Pitch x window Pitch x group Window x group	14.875 4.479 0.960 1.370 16.192 0.194	0.000*** 0.008** ns ns 0.000*** ns
vPa2	Pitch Window Group Pitch x window Pitch x group Window x group	0.667 7.607 4.449 1.725 20.071 0.031	ns 0.000*** 0.079† ns 0.000*** ns	sdPa2	Pitch Window Group Pitch x window Pitch x group Window x group	70.859 7.915 4.350 2.610 15.129 0.149	0.000*** 0.000*** 0.082† 0.064† 0.000*** ns
Pa3	Pitch Window Group Pitch x window Pitch x group Window x group	18.287 7.123 2.204 2.748 11.566 0.174	0.000*** 0.001*** ns 0.055† 0.001*** ns	sdPa3	Pitch Window Group Pitch x window Pitch x group Window x group	166.142 5.139 2.288 3.054 5.941 0.488	0.000*** 0.004** ns 0.039* 0.019* ns
Pa4	Pitch Window Group Pitch x window Pitch x group Window x group	33.673 6.407 1.514 3.868 6.326 0.168	0.000*** 0.001*** ns 0.016* 0.016* ns	sdPa4	Pitch Window Group Pitch x window Pitch x group Window x group	82.902 3.468 1.789 3.054 7.646 0.564	0.000*** 0.024* ns 0.039* 0.008** ns
Pa5	Pitch Window Group Pitch x window Pitch x group Window x group	8.496 10.896 1.836 3.080 65.663 0.254	0.006** 0.000*** ns 0.038* 0.000*** ns	sdPa5	Pitch Window Group Pitch x window Pitch x group Window x group	242.386 12.910 0.947 5.054 57.965 1.245	0.000*** 0.000*** ns 0.004** 0.000*** ns
Pa6	Pitch Window Group Pitch x window Pitch x group Window x group	0.382 6.612 0.771 1.321 7.631 0.156	ns 0.001*** ns ns 0.008** ns	sdPa6	Pitch Window Group Pitch x window Pitch x group Window x group	42.081 6.594 0.552 2.206 13.064 0.595	0.000*** 0.001*** ns ns 0.001*** ns
Pa7	Pitch Window Group Pitch x window Pitch x group Window x group	0.154 2.702 0.768 0.271 1.123 0.030	ns 0.058† ns ns ns ns	sdPa7	Pitch Window Group Pitch x window Pitch x group Window x group	10.222 1.792 0.314 0.568 7.678 0.349	0.003** ns ns ns 0.008** ns

Table 4. Mean MFDR and sd-MFDR pairwise comparisons of window sizes (where W1=method 1; W2=method 2; W3=method 3; W4=method 4) for individual /pa/ during 7-syllable train.

*Note: Comparisons significant at an overall level of 0.05 are indicated by *. Marginal significance is indicated by †.

Discussion:

There does not appear to be a significant difference in overall data from a 1000 ms analysis window to a smaller 100 ms analysis window. However, the location of the 100 ms segment does appear to alter the mean MFDR value. A greater mean MFDR was found when centered on the peak MFDR for the utterance. MFDR was found to be significantly greater at the higher fundamental frequency during the middle of a messa di voce task in the peak window analysis segment (method 3) and higher among elite singers (group A).

There is no difference in MFDR data from a 100 ms analysis segment vs. a 20 cycle analysis segment for medium low pitch (F01=330 Hz) or medium high pitch (F02=660 Hz) among professional female singers.

It is suggested that window extraction specifics be included in future research to allow closer comparison of mean MFDR.

Summary:

A moderate sized window segment appears to be sufficient for determining mean MFDR. There does not appear to be a significant advantage to using a large (1000 ms) analysis window.

Among the professional singer population, there does appear to be a difference at the glottal level in management of airflow shut-off when fundamental frequency increases among subjects who are employed in regional/national level opera companies vs. those employed at international level opera companies.

Both groups were found to increase MFDR as fundamental frequency increased.

References:

1. Titze IR (1986). Mean intraglottal pressure in vocal fold oscillation. Journal of Phonetics 14: 359-364.
2. Titze IR (1989). Physiologic and acoustic differences between male and female voices. Journal of the Acoustical Society of America 85(4): 1699-1707.
3. Holmberg EB, Hillman RE, Perkell JS (1988). Glottal airflow and transglottal air pressure measurements for male and female speakers in soft, normal and loud voice. Journal of the acoustical Society of America 84: 511-529.
4. Perkell JS, Hillman RE, Holmberg EB (1994). Group differences in measures of voice production and revised values of maximum flow declination rate. Journal of the Acoustical Society of America 96(2 Part 1): 695-698.
5. Stathopoulos ET, Sapienza C (1993). Respiratory and laryngeal function of women and men during vocal intensity variation. Journal of Speech and Hearing Research 36: 64-75.
6. Peterson KL, Verdolini-Marston K, Barkmeier JM, Hoffman HT (1994). Comparison of aerodynamic and electroglottographic parameters in evaluating clinically relevant voicing patterns. Annals of Otolaryngology, Rhinology and Laryngology 103: 335-346.
7. Sulter AM and Wit HP (1996). Glottal volume velocity waveform characteristics in subjects with and without vocal training related to gender, sound intensity, fundamental frequency, and age. Journal of the Acoustical Society of America 100(5): 3360-3373.
8. Sundberg J, Andersson M, Hultqvist C (1999). Effects of subglottal pressure variation on professional singers’ voice source. Journal of the Acoustical Society of America 105(3): 1965-1971.
9. Holmberg EB, Hillman RE, Perkell JS (1989). Glottal airflow and transglottal air pressure measurements for male and female speakers in low, normal and high pitch. Journal of Voice 3: 294-305.
10. Södersten M, Hertegård S, Hammarberg B (1995). Glottal closure, transglottal airflow, and voice quality in healthy middle-aged women. Journal of Voice 9(2): 182-197.
11. Homberg EB, Himmna RE, Perkell JS, Gress C 91994). Relationship between intraspeaker variation in aerodynamic measures of voice production and variation in SPL across repeated recordings. Journal of Speech and Hearing Research 37: 484-495.

Click here to return to the schedule/abstract listing