EmphTTS: an emphasis-control TTS with reinforcement learning

Name: EmphTTS: an emphasis-control TTS with reinforcement learning
Uploaded: 2026-10-08T11:30:21.827497+00:00
Duration: 591 s
Description: EmphTTS is a non-autoregressive text-to-speech system that uses Reinforcement Learning to control word-level emphasis. It optimizes a separate duration predictor using Group Relative Policy Optimization (GRPO) guided by an emphasis localization reward. This method successfully improves objective metrics like WER and F1, and yields strong subjective preference for controllable prosody.

EmphTTS is a non-autoregressive text-to-speech system that uses Reinforcement Learning to control word-level emphasis. It optimizes a separate duration predictor using Group Relative Policy Optimization (GRPO) guided by an emphasis localization reward. This method successfully improves objective metrics like WER and F1, and yields strong subjective preference for controllable prosody.

Open the full episode