Nasal Assimilation Across Word Boundaries in TTS

This article explores the core computational and linguistic mechanisms that govern nasal assimilation across word boundaries in connected Text-to-Speech (TTS) synthesis. It outlines how modern TTS architectures model cross-word phonological processes—commonly known as sandhi—by integrating context-aware text front-ends, prosodic boundary detection, and continuous acoustic modeling. By examining both explicit phonological rule induction and implicit deep learning representations, this guide details how synthetic speech achieves natural, human-like coarticulation during connected speech generation.

Understanding Nasal Assimilation in Connected Speech

Nasal assimilation across word boundaries occurs when a word-final nasal consonant adjusts its place of articulation to match the initial consonant of the following word. In English, for example, the alveolar nasal /n/ in "ten boys" frequently shifts to a bilabial nasal [m] ([tɛm bɔɪz]), or to a velar nasal [ŋ] in "ten cups" ([tɛŋ kʌps]). In connected, continuous speech, this regressive assimilation reduces articulatory effort and enhances fluency. For TTS engines, failing to reproduce this phenomenon results in hyper-articulated, robotic speech, while over-applying it can degrade intelligibility.

Front-End Phonological Modeling and G2P Systems

The first layer governing nasal assimilation is the text analysis front-end, specifically the Grapheme-to-Phoneme (G2P) module. Traditional and hybrid TTS architectures resolve cross-word assimilation through explicit linguistic components:

• Context-Sensitive Rewrite Rules: Rule-based systems employ formal phonological grammar (such as Chomsky-Halle rewrite rules) that scan adjacent token pairs. If an alveolar nasal precedes a labial or velar stop, the system rewrites the base phoneme to the assimilated variant before generating phonetic transcriptions.
• Joint-Token G2P Converters: Modern neural G2P systems process sliding windows or full sentence contexts rather than isolated words. By training on continuous phonemic transcriptions, neural sequence-to-sequence models automatically predict assimilated allophones based on phonetic context.
• Phonological Sandhi Modules: Dedicated sandhi parsers explicitly evaluate whether a boundary permits assimilation based on lexical category, dialectal constraints, and morphological properties.

Prosodic Boundary Conditioning

Nasal assimilation does not occur uniformly across all word boundaries; it is fundamentally gated by prosody. TTS front-ends utilize prosody predictors to determine whether two words belong to the same phonological phrase:

• Boundary Strength and Break Indices: Using standards like the Tones and Break Indices (ToBI) framework, TTS models assign break levels (typically index 0 to 4) between words. Nasal assimilation is permitted across weak boundaries (indices 0 and 1, representing clitic groups or intra-phrasal transitions) but blocked by strong boundaries (indices 3 and 4, denoting intermediate or intonational phrase breaks marked by pauses).
• Speech Rate Adaptation: Assimilation likelihood increases at faster speaking rates. Advanced TTS architectures feed speaking-rate embeddings into the front-end or duration predictors, dynamically lowering the threshold for cross-word assimilation during rapid speech generation.

Acoustic Modeling and Latent Coarticulation

In end-to-end neural TTS systems (such as Tacotron-based, FastSpeech-based, or VITS architectures), nasal assimilation is frequently handled at the acoustic modeling stage rather than through explicit phoneme substitution:

• Learned Contextual Embeddings: When characters or canonical phonemes are passed through bidirectional encoders (e.g., Conformer or Transformer blocks), self-attention mechanisms allow the representations of boundary tokens to blend. The latent representation of the word-final nasal absorbs spatial characteristics of the subsequent consonant.
• Acoustic Smoothness via Mel-Spectrogram Targets: Loss functions that penalize spectral discontinuities force the decoder to generate continuous formants. The acoustic model learns to lower the second and third formants smoothly when transitioning from a nasal to a bilabial consonant, reproducing the acoustic correlates of assimilation even if the input phoneme remains labeled as an unassimilated /n/.
• Frame-Level Alignment and Duration: Duration predictors determine the temporal space allocated to the nasal murmur. In fast transitions, the nasal murmur is shortened, and the model blends the nasalization into the preceding vowel and the closure phase of the subsequent obstruent.

Integration in Modern End-to-End Pipelines

Modern TTS pipelines achieve optimal naturalness through a balanced integration of both symbolic and continuous mechanisms. While deep neural acoustic models can learn coarticulation implicitly from large audio datasets, relying purely on implicit learning can yield inconsistent results across varying speakers and accents. Consequently, production-grade systems combine prosody-aware G2P modules—which flag eligible sandhi contexts—with high-capacity neural vocoders capable of synthesizing the subtle spectral shifts and nasalized transitions essential for realistic connected speech.