CrossAccent-TTS enables fine-grained accent control across languages

CrossAccent-TTS addresses a key challenge in cross-lingual TTS: controlling accent intensity while preserving speaker identity. Existing large language model (LLM)-based TTS systems offer strong cross-lingual generalization but provide limited control over accent characteristics. The researchers introduce a disentangled representation framework that separates speaker and accent features, enabling independent manipulation. The core innovation is the Accent Intensity Controller (AIC), which injects weighted language embeddings into the accent subspace. This allows smooth interpolation between accents and fine-grained modulation of accent strength at inference time—enabling use cases from neutralizing a non-native accent to deliberately preserving one.

The framework was evaluated on Indic Multilingual (covering phonetically diverse Indian languages) and L2-arctic (English with non-native accents) datasets. Results show CrossAccent-TTS achieves precise control of accent intensity, outperforming strong baselines in accent similarity and controllability metrics. The model maintains high speaker similarity and naturalness, even at extreme accent intensity levels. This work opens up practical applications in multilingual voice assistants, language learning tools, and audiobook production, where accent control is critical for user experience. The paper has been accepted at INTERSPEECH 2026 and is available on arXiv.