MaskGCT

MaskGCT is an open-source, research-grade text-to-speech (TTS) model developed through a collaboration between Quwan Technology and The Chinese University of Hong Kong, Shenzhen. Built on masked generative modeling and disentangled speech representation learning, it advances the state of the art in expressive, controllable, and language-agnostic speech synthesis — making it especially valuable for developers, researchers, and product teams building next-generation voice interfaces.

What it does

MaskGCT transfORMs written text into natural-sounding speech across six major languages while enabling precise voice customization. Unlike conventional TTS systems, it supports zero-shot Voice Cloning from minimal reference audio, dynamic cross-language synthesis (e.g., typing in English but speaking with a Mandarin voice), and granular control over duration, pitch, speed, and EMOtional tone — all without requiring retraining or fine-tuning for each use case.

Key capabilities

• multilingual TTS: Produces Fluent, Intelligible speech in Simplified Chinese, English, Japanese, Korean, French, and German — with consistent phonetic accuracy and prosodic naturalness across languages.
• One-shot voice cloning: Generates highly speaker-consistent audio using just SEConds of reference speech, preserving timbre, rhythm, and vocal idiOSyncrASIes.
• Prosody editing: Allows real-time adjustment of speaking rate, pause duration, pitch contour, and emotion intensity via parameterized controls — IDEal for accessibility, dubbing, or expressive narration.
• Cross-lingual voice transfer: MAIntains speaker identity while switching languages mid-utterance or across sentences, supporting seamless multilingual APPlications.
• Open architecture: Released under permissive licensing with full training code, inference scripts, and pretrained weights — facilitating reproducibility, extension, and integration into custom pipelines.

Who it is for

MaskGCT targets AI researchers exploring disentangled speech representations, ML engineers integrating robust TTS into voice agents or IVR systems, edtech and media startups building multilingual nARRation tools, and open-source contributors advancing accessible speech technology. It is not designed for non-technical end users seeking plug-and-play voice apps.

How to get started

The model is hosted publicly on GitHub within the Amphion Framework. Users need Python 3.9+, PyTorch, and CUDA-compatible hardware for local Inference. installation requires cloning the repository, installing dependencies, downloading pretrained checkpoints, and running provided inference Notebooks or CLI scripts. documentation includes usage examples for voice cloning, multilingual synthesis, and prosody manipulation — though familiarity with PyTorch and TTS fundamentals is strongly recommended.

Pricing and access caveat

MaskGCT is fully free and open-source with no usage restrictions, credits, or subscription tiers. However, it runs only via self-hosted inference — there is no official web demo, API, or cloud service. Performance, latency, and voice quality may vary depending on hardware, input text normalization, and reference audio quality. No commercial support or SLA is provided by the developers.