license: cc-by-4.0
DiaMoE-TTS: A Unified IPA-based Dialect TTS Framework with Mixture-of-Experts and Parameter-Efficient Zero-Shot Adaptation
github:DiaMoE-TTS
We utilize the Common Voice Cantonese dataset, the Emilia Mandarin dataset and dialectal data from the KeSpeech corpus and a open-source Sourthern Min dataset for training. We only release the frontend of the open-source dataset IPA here, audio data that matches the IPA frontend can be applied for or downloaded from the corresponding official link.
Short Intro
Dialect speech embodies rich cultural and linguistic diversity, yet building text-to-speech (TTS) systems for dialects remains challenging due to scarce data, inconsistent orthographies, and complex phonetic variation. To address these issues, we present DiaMoE-TTS, a unified IPA-based framework that standardizes phonetic representations and resolves grapheme-to-phoneme ambiguities. Built upon the F5-TTS architecture, the system introduces a dialect-aware Mixture-of-Experts (MoE) to model phonological differences and employs parameter-efficient adaptation with Low-Rank Adaptors (LoRA) and Conditioning Adapters for rapid transfer to new dialects. Unlike approaches dependent on large-scale or proprietary resources, DiaMoE-TTS enables scalable, open-data-driven synthesis. Experiments demonstrate natural and expressive speech generation, achieving zero-shot performance on unseen dialects and specialized domains such as Peking Opera with only a few hours of data.