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

Summary

The paper describes the bridge between the language generator and the speech synthesizer in VieCtoS, a German concept-to-speech system that produces weather reports. In concept-to-speech, the generator already knows the syntax, semantics and pragmatics of the sentence. The open problem is how to turn that knowledge into a qualitative phonological description, including stress, syllables, accents and phrasing, that a synthesizer can realize.

The pipeline works in three steps:

• A FUF-based tactical generator outputs lemmas annotated with morphosyntactic and prosodic features.
• An extended two-level phonology component processes this output. It is based on X2MORF and adds feature "filters" that let rules inspect the grammar.
• The result is a SAMPA string carrying GToBI tone and break labels.

The key technical trick is to flatten several phonological tiers onto one string, anchored to the segment skeleton. This lets standard two-level compilers handle data that is conceptually multilinear. The generator proposes only candidate accent and boundary positions. The phonology then picks or drops actual tones under contextual constraints.

The result is a working system with a transparent, modular route from information structure to intonation.

Key points

• VieCtoS runs as a pipeline: text planner → FUF tactical generator (sentence level and word level) → extended two-level phonology → phonetic interpreter → synthesizer.
• The lexicon is phonemic, written in SAMPA, and marks potential accent and syllable-boundary positions. Only inflection is handled morphotactically, while compounds and derivations are listed in the lexicon.
• Phonology runs over the whole utterance. This is because syllabification, cliticization and the intonation rules all need context larger than a single word.
• A grammar constraint defines a well-formed intonation phrase (IP): one to three pitch accents followed by an obligatory boundary tone (L-L%, L-H%, H-L% or H-H%). An L-L% boundary is triggered by assertions.
• Information structure is reduced to a few tags, such as [EMPH] and [GIVEN]. The set of GToBI contours is also deliberately trimmed to those that sound natural in synthetic speech.

Interesting ideas

• I1. Several parallel annotation tiers are collapsed into one string, with each tier's unit tied to a fixed anchor point (leftmost segment, vowel nucleus or right edge); a single-tape finite-state machine can then process data that is really multi-layered.
• I2. Context-dependent rules are given "filters" that query structured features from a higher module, so a low-level rule fires only when, say, the clause is an assertion or the word is marked unaccented.
• I3. The upstream module proposes only candidate positions (possible accents and boundaries), and a downstream component with local-context knowledge decides which ones to realize or delete.
• I4. Word behaviour can depend on the text domain: "weather" loses stress inside compounds in weather reports because the concept is already given by the text.
• I5. Valid outputs are deliberately restricted to a smaller, safer subset of what the full theory allows, since some legal forms come out unnatural in a machine setting.

For Mova

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

• I1 (tokens): Morphology, lemma and clitic-boundary layers could share one linearized token stream with explicit anchoring, so that finite-state rules stay simple and remain traceable.
• I2 (back): Surface-realization rules in generation could check frame or speech-act features through explicit filters, which keeps each choice explainable.
• I3 (back): The meaning-graph generator could mark optional slots such as pronoun versus full NP, or a comma or clause break, and let a separate realization layer resolve them using surface context.
• I4 (world): The world model could flag concepts that are given by the genre or topic, and later stages could treat them as background, for example when ranking salience or choosing referring expressions.
• I5 (back): Generation could whitelist a conservative set of constructions rather than everything English grammar permits, which reduces awkward output from the text → graph → text compressor.