DDSP-SVC is an open-source singing voice conversion project that runs on personal computers. It is built on Differentiable Digital Signal Processing and supports both real-time voice changing and offline conversion of audio files. Training and synthesis need far less hardware than SO-VITS-SVC, and training time is close to RVC. Raw DDSP output is modest, so the current version applies a rectified-flow model to lift quality, which on some datasets matches SO-VITS-SVC and RVC. The code is free to download from GitHub.
This project fits hobbyist developers, machine learning tinkerers and independent music producers who are comfortable with Python, a command line and a local GPU. The default configuration is tuned for a 44.1kHz model trained on an RTX-4060, so a mid-range gaming card is enough. Researchers comparing singing voice conversion approaches can use it as a lightweight baseline, and developers building their own voice apps can export models to ONNX. It is not a point-and-click app, and the author asks that models be trained only with legally authorized audio.
A typical workflow starts by installing the dependencies, placing a ContentVec or HubertSoft encoder, an NSF-HiFiGAN vocoder and an RMVPE pitch extractor in the pretrain folder, then preprocessing training and validation clips. After training, you can monitor progress through visualization, convert recorded vocals with the non-real-time script, or run the real-time GUI to change a voice live. Musicians might convert a demo vocal into another singer's timbre, streamers can experiment with live voice changing, and developers can batch-convert files or export the trained model for deployment elsewhere.
