TVTSyn weights: CC BY-NC 4.0 (noncommercial use with attribution). Streaming uses 100 ms model chunks; browser and network buffering add latency. Live PyTorch audio may process slower than real time on CPU.
PyTorch file conversion uses ZeroGPU (subject to its queue and quota). ONNX, speaker preparation, and live sessions use CPU. Clips are limited to 60 seconds.