MaskGCT

MaskGCT (September 2024, from the Amphion team) takes a different architectural path from most TTS: a fully non-autoregressive, two-stage masked generative codec transformer that predicts semantic tokens then acoustic tokens — no text-speech alignment supervision and no phoneme-level duration prediction required.

The Amphion codebase is MIT licensed, but note the Hugging Face model weights are tagged CC-BY-NC-4.0 (non-commercial) — a sharper restriction than the code license. For commercial use, verify the checkpoint's terms or use the code with your own trained weights. Quality is competitive with autoregressive zero-shot systems while inference is parallelizable and efficient. It is a research-grade release: excellent for researchers and builders comfortable with academic code, less polished than product-oriented TTS projects. If you want to experiment with NAR TTS architectures, this is the reference implementation.

Provider-claimed (paper): competitive with state-of-the-art zero-shot TTS on speaker similarity and naturalness — treat as provider-claimed; paper reports LibriSpeech-based evaluations. Independent third-party benchmarks are limited.