French Sandhi Rules in Text-to-Speech Systems

French phonological sandhi—primarily comprising liaison, elision, and enchaînement—modifies word-boundary pronunciation dynamically based on adjacent sounds, grammatical relationships, and register. For Text-to-Speech (TTS) systems, these phenomena transform grapheme-to-phoneme (G2P) conversion from an isolated lookup process into a complex contextual prediction task. To generate natural French audio, synthesis engines must resolve multi-word phonetic interactions across varying syntactic hierarchies, register constraints, and lexical exceptions.

The Mechanisms of French Sandhi

External sandhi refers to phonological alterations that take place across word boundaries. In French, this manifests through three primary mechanisms:

• Liaison: A normally silent latent final consonant is articulated before a following word beginning with a vowel sound or a mute h (e.g., les /le/ + amis /a.mi/ becomes /le.za.mi/).
• Enchaînement: A word-final consonant that is always pronounced resyllabifies to form the onset of the following word's initial vowel (e.g., avec elle /a.vɛ.kɛl/).
• Elision: The deletion of an unstressed final vowel—typically /ə/, /a/, or /i/—before a vowel-initial word, marked orthographically with an apostrophe (e.g., la + école becomes l'école).

While enchaînement and elision follow relatively strict phonetic and orthographic rules, liaison introduces significant ambiguity because the surface text does not explicitly indicate whether the latent consonant should surface.

Liaison Classification and Syntactic Dependencies

Liaison realization is fundamentally context-dependent and falls into three categories:

1. Mandatory (Obligatoire): Occurs within tightly knit syntactic units, such as determiners and nouns (un‿arbre), clitic pronouns and verbs (ils‿arrivent), or prenominal adjectives and nouns (petit‿enfant). Failing to pronounce these sounds robotic and grammatically incorrect.
2. Forbidden (Interdite): Blocked by specific syntactic or structural boundaries. Liaison is forbidden between a singular nominal subject and a following verb (le chat / arrive), across coordinated structures (et / elle), after the conjunction et, and before numerals like onze or huit. Realizing a forbidden liaison severely degrades intelligibility and naturalness.
3. Optional (Facultative): Found in mid-level syntactic cohesion, such as between forms of the copula être and predicates (c'est‿impossible vs. c'est / impossible) or plural nouns and post-posed adjectives (des maisons‿anciennes). Realization is governed by speaking style, tempo, and sociolinguistic register.

Because liaison behavior is dictated by syntactic depth rather than flat word adjacency, TTS pipelines relying solely on n-gram language models frequently misclassify boundary types. Accurate realization requires deep syntactic parsing to evaluate dependency relations and part-of-speech (POS) tags.

Lexical Irregularities: The "H Aspiré" Challenge

A notorious edge case in French G2P conversion is the h aspiré (aspirated h). Despite its name, this segment is phonetically silent in modern standard French, but it functions as a phonological consonant at word boundaries.

The presence of an h aspiré blocks both elision and liaison. For instance:

• Les héros (/le.e.ʁo/) contains an h aspiré, prohibiting liaison.
• Les héroïnes (/le.z‿e.ʁɔ.in/) contains an h muet (mute h), requiring liaison.

Because orthography does not distinguish between an h aspiré and an h muet, TTS systems must rely on specialized lexicon lookups. If the front-end fails to retrieve the correct lexical flag, the system will incorrectly produce /le.z‿e.ʁo/ for les héros, which native speakers perceive as an egregious error.

Prosody and Boundary Interactions

Sandhi processes are strictly bounded by prosodic structure. A liaison or enchaînement cannot cross an intonational phrase boundary or a pause.

In TTS architectures, a circular dependency often arises between the acoustic model and the phonemizer:

• The decision to insert a liaison consonant depends on whether the rhythm of the sentence permits a pause between the two words.
• The presence of a liaison consonant alters syllable counts and rhythmic duration, which in turn influences the acoustic model's prediction of phrase breaks.

If the linguistic front-end assumes a liaison will occur, but the acoustic model predicts an intonational break at that boundary, the synthesizer produces an unnatural detached consonant preceding a pause.

Architectural Approaches in TTS

Modern TTS engines address these sandhi challenges through two primary paradigms:

• Neural G2P with Contextual Embeddings: Contemporary systems use Transformer-based front-ends (e.g., fine-tuned multilingual BERT architectures) that take entire sentences as input. These models learn syntactic dependencies and can predict context-sensitive phoneme sequences, dynamically inserting liaison phonemes like /z/, /t/, or /n/ prior to feeding phonemic tokens into acoustic decoders.
• End-to-End Character-to-Waveform Models: Direct text-to-speech models skip intermediate phonetic transcription entirely, mapping characters directly to mel-spectrograms or waveforms. While this avoids explicit rule engineering, these models require vast amounts of high-fidelity training data to implicitly capture the syntactic and stylistic variance of optional liaisons, often struggling with low-frequency h aspiré words without explicit phonetic conditioning.

Achieving natural French speech synthesis requires robust front-end architectures capable of reconciling lexical exceptions, phrase-level syntactic parsing, and register consistency across word boundaries.