AgentSteerTTS: A Multi-Agent Closed-Loop Framework for Composite-Instruction Text-to-Speech

TL;DR

AgentSteerTTS introduces a multi-agent closed-loop framework that enhances expressive control in text-to-speech systems by disentangling speaker and emotion features, grounding intents with acoustic prototypes, and refining output through feedback.

Contribution

The paper presents a novel multi-agent framework for intent-faithful expressive TTS, combining disentanglement, prototype grounding, and feedback refinement for improved control.

Findings

Significant improvements over baselines on composite-instruction benchmark.

Effective disentanglement of speaker identity and emotion-prosody.

Enhanced expressiveness and fidelity in synthesized speech.

Abstract

While existing text-to-speech (TTS) models exhibit high expressiveness, fine-grained control over composite instructions remains challenging due to the structural mismatch between discrete textual intents and continuous acoustic realizations. Inspired by human cognitive decoupling, we introduce AgentSteerTTS, a multi-agent closed-loop framework designed for intent-faithful expressive control of composite instructions. First, in our framework, an adversarial disentanglement agent mitigates speaker-emotion leakage by learning separable identity and emotion-prosody subspaces with leakage-suppressing regularization. Next, a Dual-Stream Anchoring Controller grounds abstract intents using a large-scale acoustic prototype library: a Retrieval Agent selects expressive anchors, while a Synthesis Agent fuses them into continuous control vectors via gated attention. Finally, a Fast-Slow Feedback Agent refines output intensity through latent gradient correction and resolves semantic-acoustic mismatches using high-level perceptual critique. Experiments on a composite-instruction benchmark and public test sets show that AgentSteerTTS yields consistent and significant improvements to the baselines, demonstrating the effectiveness of the proposed method.