The authors propose WordVoice, a framework that transforms Large Language Model-based Text-to-Speech (TTS) from implicit generation to an explicit, highly controllable paradigm. This addresses the bottleneck of coarse-grained control in existing systems by enabling precise stylistic interventions and strict temporal alignment.
- The team constructs WordVoice-5A, a 4.7k-hour bilingual dataset with five-dimensional word-level annotations including duration, boundary, energy, pitch, and tone.
- A bound-token mechanism is introduced within the LLM to formulate an explicit "acoustic planning" process for adaptive multi-task prosodic planning.
- A fine-grained acoustic modulation module bridges the resolution gap between discrete tokens and continuous waveforms to align word-level attributes.
The framework achieves superior, decoupled control over multiple acoustic dimensions while maintaining competitive zero-shot synthesis stability.