Accent TTS Demo

Abstract

Text-to-speech (TTS) systems have advanced in producing natural speech across speakers and styles, yet zero-shot accent control remains challenging due to the entanglement of speaker and accent attributes. We propose Joycent, a novel accent TTS model that integrates both audio-based and Large Language Model (LLM)-assisted accent embeddings to achieve disentangled and natural accent synthesis. Audio-based embeddings are derived from speech using the pre-trained accent identification model WhisAID, while LLM-assisted embeddings are generated from natural language descriptions of the target accent prompt in text. Speaker embeddings are separately extracted and are conditioned together with the accent embeddings via conditional layer normalization (CLN) modules within the text encoder. Experiments demonstrate that Joycent outperforms baselines in naturalness, accentedness, and speaker similarity, showing effectiveness in zero-shot accent generation without compromising speaker identity. The audio samples are available at https://oshindow.github.io/joycent/.

Model Architecture

Figure 1. Overall model architecture.

Seen Speaker Samples (SSB0693)

Unseen Speaker Samples (G0003/SSB1340)