DiffSinger
github.com
- Free plan
- Self-hosted
What DiffSinger is
DiffSinger is a free, open-source singing voice synthesis system based on a shallow diffusion mechanism. Its maintained version adapts synthesized audio to a 44.1 kHz sampling rate, and the project includes improved acoustic models and accelerated diffusion sampling methods. Variance models and parameters let users predict and control pitch, energy, breathiness and other aspects of a vocal output. The getting-started guide covers dataset preprocessing, model training and inference on DS files using variance or acoustic models. The linked MakeDiffSinger project provides dataset-building pipelines, including recording slicing and labeling tools. DiffSinger is self-hosted, uses ONNX as its deployment format, and requires Python 3.10 or later; PyTorch 2.4.0 or later is recommended. Users prepare their own data and assets. The repository links OpenUtau and DiffScope as production or deployment projects; DiffScope is an editor powered by DiffSinger and remains under development. The project prohibits generating someone's voice without consent and includes a checkpoint utility to remove speaker embeddings when distributing models. The free plan is free.
Who it suits
DiffSinger suits members of the singing voice synthesis community and production users prepared to work with their own data and assets. Its guide covers model preparation, training and inference.
Good to know, and what to watch
- Controls pitch, energy, breathiness and other aspects.
- MakeDiffSinger provides dataset pipelines and labeling tools.
- ONNX is the deployment format.
- Commercial use is allowed.
- Requires Python 3.10 or later.
- Users prepare their own data and assets.
- DiffScope is under development.
- No free trial is listed.
Plans
Every plan and priceDiffSinger
Open source
The verdict
DiffSinger offers a self-hosted workflow for synthesizing and controlling singing vocals, with dataset tools and ONNX deployment. Users need to prepare their own assets and follow the project's consent restriction.
What the maker publishes
- Audio quality
- The maintained version adapts synthesized audio to a 44.1 kHz sampling rate instead of the original 24 kHz. source
- Control
- Variance models and parameters allow prediction and control of pitch, energy, breathiness, and other aspects. source
- Dataset tools
- The linked MakeDiffSinger project provides pipelines and tools for building DiffSinger datasets, including recording slicing and labeling tools. source
- Deployment
- The guide states that DiffSinger uses ONNX as its deployment format. source
- Editor integration
- The repository links OpenUtau and DiffScope as deployment or production projects; DiffScope is described as an editor powered by DiffSinger and is under development. source
- Installation
- The getting-started guide requires Python 3.10 or later and recommends PyTorch 2.4.0 or later. source
- Intended users
- The project says its functionality is designed for production deployment and the singing voice synthesis community. source
- Model distribution
- The getting-started guide includes a utility to remove speaker embeddings from a checkpoint for data security when distributing models. source
- Model improvements
- The project integrates improved acoustic models and diffusion sampling acceleration algorithms. source
- Purpose
- DiffSinger is a singing voice synthesis system based on a shallow diffusion mechanism. source
- Security and consent
- The repository prohibits using its functionality to generate someone's voice without that person's consent. source
- Support
- The project points users to tutorials, GitHub issues and discussions, and QQ and Discord communication groups. source
- Training and inference
- The guide describes preprocessing datasets, training models, and running inference on DS files using variance or acoustic models. source
Checked against the maker's pages on 4 Oct 2026.