Skip to content
← Models
Open weightsCUHK-Shenzhen / Shanghai AI Laboratory (OpenMMLab)/10 trustVerified

Amphion

The open toolkit for anything-to-audio: TTS, singing voice conversion and synthesis, text-to-audio and vocoders in one MIT-licensed framework with 10,000+ stars.

Visit official page →

About

Amphion (arXiv 2312.09911, December 2023; v0.2 January 2025) is the OpenMMLab toolkit for audio, music and speech generation research, built at CUHK-Shenzhen and the Shanghai AI Laboratory under Zhizheng Wu. Its stated goal is converting any input into audio: TTS (including MaskGCT and zero-shot Vevo), singing voice conversion (DiffSinger, VISinger-D recipes), singing voice synthesis (diffusion-based, supported in v0.2), text-to-audio, text-to-music, plus a vocoder zoo and standardized evaluation metrics. Everything ships with data-preprocessing pipelines, pre-trained models on Hugging Face, and the visualizations the project is known for. At over 10,000 GitHub stars with MIT licensing, it is the most-adopted open framework in this batch. The honest caveat: it is a research toolkit, not a product. Running the recipes still requires GPU time and ML comfort.

Specs

Context window

Model-dependent; toolkit includes recipes from short utterances to song-length SVC (no single published limit).

Pricing

Free / open source: MIT

Free quota

Fully open: code, pre-trained models and datasets on GitHub and Hugging Face; self-host

API access

No hosted API: self-host via the toolkit; pre-trained models on Hugging Face

Weights

Open weights

Released

Not disclosed

Benchmarks

  • No single Amphion benchmark; the toolkit reproduces published results of its component models (DiffSinger, MaskGCT, Vevo)
  • Evaluation-metric suite ships with the repo for consistent comparisons

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

singing-voice-conversionsinging-voice-synthesistext-to-speechtext-to-audiovocoder-benchmarkingresearch-framework

Trust score — 9/10

/10 trustVerified

Scored 9/10 (TIER verified, metrics-v1). Evidence: peer-reviewed overview paper; 10,000+ GitHub stars; MIT license; pre-trained models on Hugging Face; v0.2 shipped January 2025 with active maintenance. Strengths: the broadest, best-organized open toolkit for singing and audio generation, with reproducible recipes and its own evaluation metrics. Docked for: (1) it is a framework, not a single model, so quality claims belong to its components; (2) no Amphion-level benchmarks exist beyond component reproductions. License: MIT.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
9.0/10
Maintenance · recency, releases, issue load
9.0/10
Openness · license + docs
10.0/10
Manual review · sanity read + red-flag check
9.0/10
  • · arXiv 2312.09911, December 2023; CUHK-Shenzhen / Shanghai AI Lab, OpenMMLab
  • · 10,000+ GitHub stars; v0.2 shipped January 2025
  • · SVC supported; SVS supported in v0.2; pre-trained models on Hugging Face
  • · Framework-level entry: quality claims belong to component models; no logo asset found, logoUrl omitted

Sources