Table 1.
Illustration of the poem modification on the first stanza of William Blake “Ah Sun-flower!”.
The glossary column provides more detailed explanations of the specific modifications. The last column gives average number of syllables of the original and modified versions of the selected 40 poems.
Fig 1.
Illustration of the acoustic analysis workflow.
A. Top: The digitized speech signal was annotated, using syllabic units (example poem: August von Platen [1814], Lass tief in dir mich lesen). Bottom: The pitch contour was obtained by a two-pass fundamental frequency (F0) estimation. From sonorous parts, the mean pitch at three measurement positions was calculated. B. Mean pitch values were mapped onto semitones, using the MIDI convention, and syllable duration was mapped onto musical length. For illustration purposes, the resulting notation was shifted two octaves up. C. Discrete pitch and duration values were subjected to autocorrelation analyses. Apart from an overall measure of autocorrelation strength, the study focused on autocorrelation values at lags that correspond to poetic structure, such as (all) individual lines, rhyming lines, and stanzas.
Table 2.
Basic acoustic properties of the recitations by the professional speaker, synthetic voices, and naïve (nonprofessional) speakers.
Each speaker in the control group of 10 speakers produced a subset of 16 poems (8 original [A] versions, 8 modified [E] versions). The syllable rate is computed as the number of non-silent syllabic units per time unit.
Fig 2.
Mean melodiousness ratings for all poem versions.
Notably, melodiousness ratings for the original poems (version A) did not differ between experiments.
Fig 3.
Correlations of melodiousness ratings with stanza-lag autocorrelations, based on pitch (red) and on duration (blue), for the 40 original poems, recited by the professional speaker.
Pitch-based autocorrelations were significantly correlated with melodiousness ratings.
Fig 4.
Illustration of textual unit and poem version effects on pitch- and duration-based autocorrelations.
A. The scaling of autocorrelations with textual unit (all lines<rhyming lines only<stanzas) crucially depended on acoustic dimension and on speaker (PS: professional speaker, SV: synthetic voice). B. Illustration of the poem version effect for the professional speaker. The strongest effect is seen for pitch-based autocorrelations across stanzas. Error bars indicate standard errors of the mean.
Fig 5.
Modification effect on autocorrelations at the textual unit of all verse lines (blue), rhyming lines only (red) and stanzas (green), obtained from renditions by nonprofessional speakers (N = 10).
Overall, autocorrelations are higher for original than for modified poems, but differ depending on acoustic dimension (pitch or duration) and textual unit (all lines, rhyming lines only, stanzas). Error bars indicate standard errors of the mean.
Fig 6.
Overall correlations between melodiousness ratings and mean autocorrelations at stanza lag, plotted separately for pitch-based autocorrelations (red) and duration-based autocorrelations (blue).
These correlations involve all poem versions.
Fig 7.
Dependency of pitch autocorrelations and musical settings.
Autocorrelation values are higher for poems that have been set to music than for poems that have not been set to music. This particularly holds for autocorrelations across stanzas. Error bars indicate the standard error of the mean.