CMsmartvoice/One-Shot-Voice-Cloning

Combines a U-Net architecture with Adaptive Instance Normalization (AdaIN) layers to enable robust speaker and style transfer from a single reference audio sample, automatically estimating duration statistics without manual annotation. Built on TensorFlowTTS, it uses a three-stage pipeline (duration model, acoustic model, vocoder) trained exclusively on neutral speech corpus to synthesize arbitrary text in cloned voices. Supports both Python inference and Google Colab notebooks, with pre-trained models available for immediate use.

245 stars. No commits in the last 6 months.