EMNLP 2026 · Main Conference

LoopTTS

Diagnose, Then Refine: A Closed-Loop TTS System with AudioLLM-Guided Correction

LoopTTS is a closed-loop TTS system that uses an AudioLLM Judge to diagnose local prosodic defects and guide expressive re-synthesis.

Audio Demo Paper Soon Repository

Filter · Judge · Refiner

Given an initial utterance, LoopTTS filters severe failures, identifies salient word-level prosodic issues, and produces structured instructions covering global emotion, speed, and pitch together with local stress and pause cues. A purpose-trained Refiner then performs guided re-synthesis conditioned on the original audio, target text, and instruction.

DemoInteractive audio comparisonsAvailable

TrainingRefiner training frameworkComing soon

Refiner-DBProsody refinement datasetComing soon

Audio Samples

Fine-grained prosody refinement

Listen to global expressive control, local stress and pause control, and Judge-guided before/after refinement.

Loading samples…