wtf is viitor-voice-nar?

ViiTorVoice-NAR is a Python-based speech generation system focused on voice cloning and audio editing. NAR stands for non-autoregressive, a technical term describing how the system generates audio: instead of producing one segment at a time in sequence, it can work on multiple parts of the output in parallel, which is one of the reasons it achieves low latency. The README states a first-frame response time of around 60 milliseconds. The system has three main capabilities. Voice cloning takes a short audio clip of a person speaking and synthesizes new speech in that voice for whatever text you provide. Local editing is more precise: you give it a source audio file, the original transcript of what was said, and an edited version of the text, and the system identifies exactly which words changed and re-synthesizes only those sections, leaving the rest of the audio untouched. The third capability is emotion and style control: you can insert special tags into the text (for example, marking a sentence as happy or surprised), and the system uses those tags to shape the tone of the generated speech. The architecture runs as a set of separate background services that communicate over gRPC, a high-speed network protocol designed for internal service communication. An HTTP gateway sits in front of them so you can make simple web requests to trigger cloning or editing without needing to understand gRPC. A management script starts, stops, and checks the status of all services at once. Setup requires running an initialization script that creates a Python virtual environment and installs dependencies, then downloading the model files from Hugging Face (about 4GB or more, though the README does not state an exact size) into a local directory. The HTTP service then listens on port 7861, and the README provides ready-to-run curl command examples for each feature. A live demo is available on Hugging Face Spaces, and the model weights are published at the linked Hugging Face model page. The README acknowledges OmniVoice and DualCodec as architectural inspirations.