Four layers and a third of the truth: what jitter measures, and why it stops measuring it

A recording of a sustained vowel is not one thing. Sorting each cell of its time-frequency plane by how many glottal cycles the component repeats itself unchanged decomposes it into four layers whose powers sum exactly to the total. Across 465 speakers with an electroglottographic clock the median shares are 86.2 % persistent, 9.9 % slowly lost, 2.4 % quickly lost and 0.7 % non-repeating. JITTER READS ONE OF THEM. Against jitter measured from the electroglottographic channel, the non-repeating layer correlates at Spearman rho = +0.69 and the persistent layer at -0.44. The measure the field has used for sixty years to describe a voice is a description of under one per cent of that voice's power. THE RESPONSE FUNCTION. On 8,191 recordings across five vowel and pitch conditions with the same truth, the least-squares slope of acoustic jitter on true jitter is 0.296 (95 % CI 0.270-0.324): about three tenths of a real change reaches the printed number. The measured value lies below the truth in 90 % of recordings, median ratio 0.750. All four jitter definitions in common use -- local, RAP, PPQ5, DDP -- give the same slope to within 0.05, so this is not a quirk of one formula. THE RESPONSE COLLAPSES IN CLEAN RECORDINGS. Split by harmonics-to-noise ratio, the slope falls monotonically from 0.422 below 10 dB to 0.012 (95 % CI 0.002-0.028) above 32 dB. The regime the field trusts most is the one in which the number carries least information about the larynx. This is NOT range restriction: the standard deviation of the true value is higher in the cleanest band than in the band below it while the slope is six times lower, and the rank correlation falls as well. Eighty-five per cent of this corpus sits at 20 dB or above, where the slope is 0.149 or less. SHIMMER BEHAVES DIFFERENTLY. Measured the same way on the same recordings, its pooled slope is 0.822 and it does not collapse. Whatever is wrong with jitter is specific to jitter, and the two should not be reported as equally trustworthy. BIAS AND UNRELIABILITY ARE SEPARATE DEFECTS with opposite remedies, and the paper reports them separately. Averaging shrinks the scatter and leaves the systematic three tenths untouched. An interpretation is offered for why the collapse happens -- the layer jitter reads shrinks as recordings get cleaner -- together with the test that would settle it, because the two analyses ran on different subsets and cannot be joined speaker by speaker with the data available. The per-recording derived tables and the analysis script are included. The recordings belong to the Saarbruecken Voice Database and are not redistributed.