WordVoice: Explicit and Decoupled Multi-Dimensional Word-Level Control for LLM-Based TTS

TL;DR

This paper proposes a framework for explicit and decoupled control over word-level acoustic attributes (duration, boundary, energy, pitch, tone) in LLM-based TTS.

Problem

State-of-the-art LLM-based TTS systems lack the fine-grained, word-level control necessary for precise stylistic interventions and strict temporal alignment required in applications like audiobook narration and dubbing. This bottleneck is due to the scarcity of fine-grained annotated datasets and the architectural challenge of integrating multi-dimensional control signals into discrete autoregressive generation.

Approach

Results & Contribution

Extensive experiments show that WordVoice achieves superior, decoupled control over multiple acoustic dimensions while maintaining competitive zero-shot synthesis stability. The key contributions are the release of a large-scale word-level annotated dataset and the proposal of a new paradigm for precise control in LLM TTS.