DDSP-SVC
The lightweight alternative to RVC and SO-VITS-SVC: singing voice conversion built on differentiable DSP, with training that runs on a personal computer and a real-time voice-changer GUI.
About
DDSP-SVC (github.com/yxlllc/ddsp-svc, March 2023; MIT) is an open singing voice conversion system built on DDSP, differentiable digital signal processing, instead of the heavy diffusion or VITS stacks its rivals use. The pitch is simple: much lower hardware requirements for training and synthesis, training speed close to RVC, and a real-time voice-changing GUI that runs on an ordinary PC. Raw DDSP output is admittedly rough; the project compensates with a quality-enhancement stage that evolved from a pre-trained vocoder enhancer to a shallow diffusion model and now a rectified-flow model, which the author says can match SO-VITS-SVC and RVC quality on good datasets (provider-claimed). It supports ContentVec content features, RMVPE/FCPE pitch extraction, timbre mixing between speakers, and key shifting. With around 2,700 stars and commits still landing, it is one of the most actively maintained open SVC projects. The repo carries a clear disclaimer: train only on legally obtained, authorized data.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Quality parity with SO-VITS-SVC and RVC on some datasets after the rectified-flow enhancer (provider-claimed; no controlled third-party benchmark found)
- No published MOS or objective metrics; evaluation is community A/B listening
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: MIT-licensed, around 2,700 GitHub stars, and commits within the last two weeks, genuinely alive. The real-time GUI and low training cost are verifiable from the repo, not just claimed. Strengths: the best-documented low-resource path to singing voice conversion, honest disclaimers about data legality, and an active community producing compatible models. Docked for: (1) no published objective or MOS benchmarks, so quality claims are provider- and community-claimed; (2) a one-maintainer community project, so the bus factor is low; (3) results depend heavily on the training dataset, and weak datasets sound weak. License: MIT.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · yxlllc/ddsp-svc, created 2023-03-04, ~2,700 stars, commits within last 2 weeks
- · DDSP-based SVC; rectified-flow enhancer for quality (provider-claimed parity with RVC/SO-VITS-SVC)
- · Real-time GUI voice changer; timbre mixing, key shift
- · No published objective/MOS benchmarks; single-maintainer project