5 Jitter (Frequency Perturbation)
Living draft. This chapter reflects the sources cited below and will be revised as further sources are incorporated. Found an error? Use the Report an issue link in the sidebar.
5.1 Definition
Jitter is the variability of the fundamental frequency — or, reciprocally, of the fundamental period — from one cycle to the next. It measures how much a given period differs from the period that immediately follows it, not how much it differs from a cycle at the other end of the utterance. It therefore captures the frequency variability that is not accounted for by voluntary changes in F0 (Baken and Orlikoff 2000).
In an ideal and perfectly stable mechanism there would be no difference between fundamental periods except when a speaker purposely changed pitch, and perturbation would be zero. To the extent that jitter is not zero, perturbation is an acoustic correlate of erratic vibratory patterns arising from diminished neuromotor and aerodynamic control and from altered tissue rheology (Baken and Orlikoff 2000).
5.2 Computation
There is a very large number of ways an index of frequency perturbation might be calculated, and each has different advantages and drawbacks. All of them can legitimately be called indices of jitter. The plurality of methods and the perils of interpretation call for judicious — “very judicious” — application (Baken and Orlikoff 2000).
Absolute measures ignore the speaker’s F0. Several have been devised, although none has proven very popular. Lieberman’s perturbation factor is the percentage of period differences equal to or greater than half a millisecond. The directional perturbation factor (DPF) is the percentage of period differences that change sign (Baken and Orlikoff 2000).
F0-related measures exist because the magnitude of perturbation correlates considerably with mean fundamental frequency: larger cycle-to-cycle period differences go with longer fundamental periods, so higher frequencies tend to show less perturbation. In one series of six men sustaining tones at roughly 2-semitone intervals, mean perturbation decreased as F0 increased with a correlation of −.95, though the relationship is neither monotonic nor invariable. There is no way to compensate exactly for mean F0 and obtain an “uninfluenced” index; the best compromise is a ratio of mean perturbation to mean period, which is what most jitter indices compute (Baken and Orlikoff 2000).
Relative average perturbation (RAP), also called the frequency perturbation quotient, addresses a problem all the preceding indices share: even during a sustained vowel, F0 may drift slowly in ways that are not of interest but that inflate the measure. RAP estimates a drift-free value for each period by averaging it with the periods before and after it — a three-point straight-line average (Baken and Orlikoff 2000).
5.3 Normative data
This reference publishes no jitter cutoff.
No source reviewed here states a Praat version and settings for a jitter reference value. Under the rule set out in Pipeline Dependence and Reference Values, an algorithm-defined measure can be compared with a reference only when the reference states the same algorithm, settings, software version and task. No published jitter reference meets that test against a Praat-based pipeline, so the absence is declared rather than filled.
| Population | Task | Pipeline | Cutoff | Class |
|---|---|---|---|---|
| Any, measured in Praat | Sustained /a/ | — | No published reference | — |
There is a second reason for caution, independent of the first and specific to jitter: at normal speaking frequencies the measurement error is of the same order as the quantity being measured. See Confounds and cautions below.
The values below are from other pipelines and are shown for context only. Each is attributed to the study the textbook names; none states a software version, and where the analysis system is named it is given.
| Measure | Population | Task | Value | Pipeline | Class |
|---|---|---|---|---|---|
| Jitter ratio | 6 men, 26–33 y | Sustained vowel, mean F0 109.8 Hz | 0.461 (range 0.408–0.590); absolute 0.042 ms | Not stated | unclassified |
| Jitter ratio | 6 men, 68–80 y | Sustained vowel, mean F0 120.1 Hz | 0.625 (range 0.468–0.740); absolute 0.053 ms | Not stated | unclassified |
| Jitter factor | 4 men, 21–37 y | Sustained vowel, four frequencies | 0.47 at 102 Hz; 0.53 at 142 Hz; 0.43 at 198 Hz; 0.97 at 275 Hz | Not stated | unclassified |
| Jitter factor | 5 men, 55–71 y | Sustained vowel, mean F0 115.3 Hz | 0.99 (range 0.76–1.49) | Not stated | unclassified |
| RAP × 100 | 24 men, 18–25 y | Sustained /a/, mean F0 117.8 Hz | 0.38 (SD 0.169) | Kay Elemetrics Visi-Pitch; version and settings not stated | unclassified |
| RAP × 100 | 25 women, 18–25 y | Sustained /a/, mean F0 222.9 Hz | 0.89 (SD 0.622) | As above | unclassified |
| RAP × 100 | 50 men, mean 30 y | Sustained /a/, mean F0 107.5 Hz | 0.28 (SD 0.12) | Not stated | unclassified |
| DPF | 20 men, 20 women | Sustained /a/, /u/, /i/ | Men 46.24, 49.26, 46.37; women 48.79, 52.77, 52.04 per cent sign change | Not stated; called tentative norms | unclassified |
| Relative jitter | 14 women per group, three age groups | Sustained vowel, onset and offset discarded | Young 0.69, middle-aged 0.57, elderly 0.66 per cent; groups did not differ significantly | Kay Elemetrics CSL 4300, voice analysis function; 51.2 kHz, 2-second frame; MiniDisc at 44.1 kHz | unclassified for a Praat pipeline |
The last row of Table 5.2 is the best-documented in it and still cannot serve as a cutoff for a Praat measurement. Its pipeline is stated completely; it simply belongs to another engine.
5.4 Confounds and cautions
The measurement error can be as large as the measurement. Perturbation can be no more accurate than the period determinations it is computed from, which are limited by the sampling frequency and the F0-extraction method. At a sampling rate of 25,000 per second the maximum relative jitter error is about 0.2 per cent for a voice averaging 100 Hz, and about 0.6 per cent at 300 Hz. Compared with the expected jitter of a normal voice, the error in both cases is quite large, and raising the sampling rate enough is generally not feasible (Baken and Orlikoff 2000).
Never report an index without inspecting the F0 record. “Undertaking a numerical analysis of vocal frequency perturbation without checking for signs of patterns of F0 variation is very poor practice indeed.” The textbook’s illustration is an 81-year-old woman diagnosed with spasmodic dysphonia whose F0 record shows a slow frequency tremor with a superimposed faster tremor and intervals of diplophonia. A single jitter index would have missed the salient features of her problem (Baken and Orlikoff 2000).
Extraction algorithm. There are three basic choices — zero crossing, peak picking and waveform matching. Peak-picking and zero-crossing methods cannot be relied upon for adequate precision, especially for the perturbations characteristic of normal voices, and are seriously affected by additive noise. Waveform matching loses accuracy when frequency variations are large. Filtering can improve the first two in principle, but may also degrade the resulting measure (Baken and Orlikoff 2000).
Vibrato is not jitter. Vibrato is a periodic variation of F0 of up to about 3 per cent, or about 1.4 semitones, at about 4 to 7 cycles per second, usually accompanied by comparable amplitude variation (Baken and Orlikoff 2000).
Minimum sample. For normal voices with only minimal roughness, as few as 30 consecutive cycles may be adequate, though more cycles reduce the variability introduced by analysing from different starting points, which carry different amounts of residual onset transition (Baken and Orlikoff 2000).
Recording. Analog tape recorders introduce considerable frequency and amplitude variation of their own, which inflates perturbation measures; they should not be used unless absolutely necessary. A digital recorder adds significantly less, and acquiring the sample through the A/D system directly into the computer is preferable. Cardioid condenser microphones with a balanced output are preferred, set close to the speaker (Baken and Orlikoff 2000).
Signal typing. Numerical jitter is reliable only on Type 1 signals. On Type 2 signals, visual analysis of stable segments is the fallback; on Type 3 signals neither numerical nor visual analysis holds (Titze 1995; Behlau, n.d.). Taken with the finding that measures of aperiodicity apparently cannot be reliably applied to voices that are even mildly aperiodic (Bielamowicz et al. 1996), this confines reliable jitter to Type 1 signals — which excludes much of the clinical population the measure is used on.
System dependence. Different analysis systems may produce dissimilar results on the same signal, and software characteristics do not account for all of the variability; until inter-laboratory standards are agreed upon, comparison of data sets must be undertaken with caution (Baken and Orlikoff 2000). The size of the problem is set out in Pipeline Dependence and Reference Values, which compares three and four analysis systems directly (Karnell et al. 1995; Bielamowicz et al. 1996).
5.5 Points of disagreement
Direction of the F0 effect. The sources disagree on the sign. Baken and Orlikoff state that larger cycle-to-cycle differences go with longer fundamental periods, so perturbation decreases as F0 rises (Baken and Orlikoff 2000). The Behlau chapter states that low F0 and weak intensity give lower values (Behlau, n.d.). These cannot both hold; this chapter follows the English primary, which reports the correlation and the caution that the relationship is not monotonic.
Minimum number of cycles. Baken and Orlikoff put the minimum at about 30 consecutive cycles for a normal voice (Baken and Orlikoff 2000); the Behlau chapter recommends at least 1 second and at least 100 cycles in its F0 guidance (Behlau, n.d.). The two are not strictly comparable — one is a minimum for a valid perturbation estimate, the other a recommendation for F0 analysis — but a clinician reading both will meet different numbers.
The figure 0.5 means two different things. In Baken and Orlikoff, 0.5 ms is the threshold inside Lieberman’s perturbation factor (Baken and Orlikoff 2000). In the Behlau chapter, 0.5 per cent is given as an upper limit of normality for relative measures, with no software, version or settings attached (Behlau, n.d.). They are not the same quantity, and the second has no traceable pipeline, so it is not listed as a reference value above.
5.6 Compute it in PhonaLab
PhonaLab is developed by the author of this reference; see Competing interests.
PhonaLab reports jitter with its algorithm, settings and task stated alongside the value. It deliberately shows no normal band and no severity band for jitter, because no published reference value is compatible with its pipeline — the reasoning is in Table 5.1 above.