Pre-trained VITS checkpoint for synthetic wakeword generation

Pre-trained VITS checkpoint for synthetic wakeword generation, converted from piper-sample-generator (https://github.com/dscripka/piper-sample-generator) v1.0.0.

Converted from pickle to state_dict format (torch.load(..., weights_only=True)). weight_norm removed from dec (Generator) and flow layers before saving.

Files:

• en-us-libritts-high.pt — VITS state_dict, 904 speakers, 22050 Hz (166 MB)
• en-us-libritts-high.json — Model config with synthesizer kwargs and phoneme ID map (21 KB)

Downloaded automatically by livekit-wakeword download.

Original model: rhasspy/piper-sample-generator v1.0.0 (MIT license)