📄 From Information Structure to Intonation: A Phonological Interface for Concept-to-Speech

Summary

The paper presents the phonological interface inside VieCtoS, a German concept-to-speech system. The interface sits between a FUF-based tactical generator and a speech synthesizer. In concept-to-speech, syntactic, semantic and pragmatic information is already known, so it does not have to be guessed from text. That information still has to be turned into both segmental and prosodic descriptions.

The authors extend two-level morphology with feature-based "filters", which let phonological rules consult grammatical information. They use it as a two-level phonology over a lexicon of SAMPA phonemic forms. To keep several phonological tiers in one string, they linearize them: syllables, stress, accents and boundary tones are attached to a segmental skeleton by simple linking conventions. This lets standard two-level compilers process the data.

For intonation, the generator only marks candidate pitch-accent and phrase-boundary positions. The two-level rules then map these to GToBI labels or delete them, based on information-structure features and phonological context.

The system works and produces German weather reports. It shows that a lean finite-state formalism can handle prosody that spans several phonological dimensions.

Key points

• Pipeline: text planner → FUF tactical generator (sentence level and word level) → extended two-level phonology → phonetic interpreter → synthesizer.
• All linguistically motivated transformations happen in the two-level component, and none in the synthesizer.
• The lexicon holds SAMPA phonemic entries with potential accent and syllable-boundary marks.
  • Only inflection is handled morphotactically.
  • Compounds and derivations are stored in the lexicon.
• Processing covers the whole utterance, because the intonation rules need the widest context.
• Syllabification, segments and word stress are treated together, since they depend on each other.
  • Example: German "Wetter" loses its stress inside compounds in weather reports, where the concept is textually exophoric.
• Intonation phrases are constrained to 1–3 pitch accents plus an obligatory boundary tone, using GToBI labels.
  • A rule licenses the L-H% boundary after a falling accent.
  • After a rising accent, L-H% is licensed only if at least two unaccented syllables follow.
• Only a few conceptual tags, such as [EMPH] and [GIVEN], drive accentuation. The set of GToBI contours is deliberately reduced to a subset that sounds natural in synthetic speech.

Interesting ideas

• I1. Several annotation tiers are flattened into one string by tying each unit to a segment skeleton with a fixed anchor (left edge, nucleus or right edge), so ordinary single-string finite-state tools still work.
• I2. The upstream module only proposes underspecified candidates (abstract accent T, boundary B), and a downstream constraint stage either realizes each one concretely or deletes it.
• I3. Local rules are allowed to query high-level features through explicit filters (sentence type, givenness, textual exophora) instead of mixing heuristics together.
• I4. The authors restrict output to a conservative subset of contours that a theory allows, because some forms that humans produce sound wrong when a machine produces them.
• I5. Some rules check whether the phonological context leaves enough room for a pattern, such as enough syllables to realize a fall-rise.

For Mova

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

• I1 (tokens): extra layers such as clitic boundaries, morpheme splits or sentence-stress marks could be anchored to token positions the same way, keeping rules for the en crate simple and inspectable.
• I2 (back): generation from the meaning graph could insert optional slots (pronoun vs. full name, comma, emphasis) and let a separate surface-constraint pass decide, which keeps each decision explainable.
• I3 (frames): pragmatic decisions such as speech-act type or givenness could be exposed as named features that later rules test explicitly, so every effect traces back to a stated condition.
• I4 (back): generated text could be limited to a whitelist of constructions known to read naturally, trading expressiveness for reliability in the text → graph → text compressor.
• I5 (back): realization rules could check whether there is enough material before choosing a construction, for example avoiding a cleft or a heavy fronting when the constituent is too short.