plachtaa/seed-vc

Seed-VC is a voice conversion tool that can take a recording of someone speaking and re-synthesize it to sound like a different person's voice, all without requiring any training on the target voice in advance. You provide a short audio clip of the reference voice (anywhere from one second to thirty seconds), and the model uses that to convert the speech from your source recording into the target speaker's voice. This is called zero-shot voice conversion, meaning the system works on voices it has never seen during training. The tool supports three main use cases. The first is standard speech voice conversion, where a recorded spoken audio file is converted to match a reference voice. The second is real-time voice conversion, which processes audio with roughly 400 milliseconds of total delay, making it usable for live scenarios like online gaming, meetings, or streaming. The third is singing voice conversion, which applies the same idea to singing rather than speech and includes controls for pitch adjustment and key shifting. Four model variants are available, ranging from a 25-million-parameter model optimized for real-time use to a 200-million-parameter model designed for highest-quality singing conversion. A newer v2 model also includes accent and speaking style transfer on top of voice timbre matching. All model weights download automatically on first use from Hugging Face, a platform for hosting AI model files. For those who want better performance on a specific speaker, the repository supports fine-tuning the model on custom recordings. The bar for this is low: a minimum of one audio clip per speaker and about two minutes of GPU training time are enough to start. Usage is through a command-line Python script or a web-based graphical interface built with Gradio. A live demo is also available on Hugging Face Spaces. The project targets Python 3.10 on Windows, Linux, and Mac with Apple Silicon chips.