CrossAccent-TTS: Cross-Lingual Accent-Intensity Controllable Text-to-Speech via Disentangled Speaker and Accent Representations

Researchers introduce CrossAccent-TTS, a machine learning framework that enables precise control over accent characteristics in cross-lingual text-to-speech systems. The technology uses an Accent Intensity Controller to allow smooth interpolation between accents while maintaining speaker identity, with particular applications for low-resource Indic languages.

CrossAccent-TTS addresses a specific but important gap in speech synthesis technology: the ability to control accent characteristics in multilingual TTS systems without sacrificing speaker identity or naturalness. Traditional LLM-based TTS models excel at cross-lingual generalization but lack fine-grained control mechanisms for accent manipulation, a limitation that constrains practical applications in diverse linguistic markets.

The framework's innovation centers on disentangling speaker representations from accent representations through an Accent Intensity Controller that uses weighted language embeddings. This technical approach enables inference-time control—users can smoothly blend between accent profiles and adjust accent strength independently, solving a problem particularly acute for Indic languages which have limited training data but high phonetic diversity.

For the speech synthesis and localization industries, this work has tangible implications. Content creators, voice production teams, and multilingual platform developers gain tools for more nuanced voice generation without requiring multiple speaker recordings. The technology's demonstrated performance on Indic languages signals growing AI investment in underserved linguistic markets, where previous solutions were either unavailable or produced lower quality outputs.

The research indicates broader momentum toward interpretable, controllable AI systems rather than black-box models. As companies deploy multilingual services globally, accent control becomes commercially relevant—enabling authentic regional representation in audiobooks, games, virtual assistants, and customer service platforms. The work's focus on speaker similarity preservation during accent conversion suggests maturation in speech synthesis, moving beyond basic intelligibility toward production-quality outputs.

• →CrossAccent-TTS enables precise accent intensity control in cross-lingual speech synthesis while preserving speaker identity and naturalness
• →The Accent Intensity Controller uses weighted language embeddings to allow smooth accent interpolation at inference time without retraining
• →Framework demonstrates significant performance improvements on Indic Multilingual and L2-arctic datasets compared to existing baselines
• →Technology addresses critical gaps in low-resource and phonetically diverse language synthesis, expanding AI accessibility beyond high-resource languages
• →Development signals commercial demand for controllable, nuanced voice generation in multilingual platforms and content creation