📄 Study and quantification of the declination for the Arabic speech synthesis system PARADIS.

Summary

The paper investigates declination in Arabic: the gradual fall of the fundamental frequency (F0) across an utterance. The goal is a melody model for PARADIS, a TD-PSOLA text-to-speech system that concatenates disyllables. The model predicts a declination line and then lays accent-related local events on top of it.

The authors built an analysis tool, WinF0, which does three things:

• estimates F0 with a normalized AMDF-style dissimilarity measure,
• supports semi-automatic syllable labeling,
• fits top and base lines by least-squares regression over local F0 maxima and minima.

Slopes are measured in semitones per syllable. Three male speakers recorded 90 isolated sentences covering the Arabic sentence modalities: declarative, exclamatory, imperative, call and interrogative. For each modality, slope is expressed as a power function of syllable count, and the per-speaker fits are averaged into a speaker-independent curve.

Declination is confirmed for Arabic, and its strength differs by modality. A small study of F0 reset at pauses matches earlier findings for French. The resulting curves let a synthesizer derive a natural-sounding F0 envelope from two inputs only: sentence type and length.

Key points

• Corpus and method: 90 isolated sentences from 3 male speakers. Declarative and exclamatory sentences were read in a narrative style; imperative, interrogative and call sentences in a dialogue style.
• Units: the pitch interval is computed as 12·log2(f1/f2) semitones and then divided by the number of syllables.
• Declarative modality: all slopes are negative and become flatter per syllable as sentences get longer. The averaged fits are:
  • base line: Y = −12.04·X^−1.50
  • top line: Y = −13.65·X^−1.36
• Questions are split by particle:
  • Wh-particles (man, mA, limAvA, kayfa, …) give a falling final contour.
  • hal, ea, or no particle gives a rising final contour. Here the top line rises (average coefficient +198.44, exponent −2.76), while the base line still falls.
• F0 reset was analysed only where it coincides with a pause. When a falling-final sentence is split, the non-final groups end rising; a rising-final sentence shows the reverse pattern. Both observations agree with Martin (1987) on French.

Interesting ideas

• I1. Modelling a per-unit rate as a power function of total length (a·X^b) turns a fuzzy trend into two interpretable parameters per category, and the R² of the fit shows how well the law holds.
• I2. Each speaker gets their own fitted curve, and the curves are then combined by geometric mean into a speaker-independent model, so no single voice dominates the result.
• I3. Surface particles decide in advance which melodic class a question belongs to: wh-words predict a falling contour, while hal/ea or no particle predicts a rising one. A cheap lexical cue therefore stands in for a hidden property.
• I4. The dissimilarity measure is normalized so that one absolute, experimentally set threshold can make the voiced/unvoiced decision across all frames.
• I5. Envelope lines are fitted only to local extrema rather than to every sample, which captures the global trend while ignoring local accent bumps.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): when evaluating parsers or readers, fit error rate against sentence length per construction type instead of reporting one average, so length effects become explicit and comparable.
• I2 (learn): per-witness parser statistics (for example, attachment-confidence curves) could be fitted separately and then pooled the same way before building silver data.
• I3 (frames): speech-act tagging could likewise split questions into wh- versus polar classes from overt markers, giving the answer stage a different expected answer shape for each.
• I4 (learn): normalizing agreement scores between parsers before thresholding would let a single fixed cutoff decide accept versus reject for silver trees of any length.
• I5 (back): when generating text from the meaning graph, discourse-level properties such as topic decay could be tracked from salient peak mentions only, which keeps local detail out of the global model.