ViiTorVoice-NAR

Editing one word inside finished narration normally means regenerating the whole take; the non-autoregressive design of ViiTorVoice-NAR makes the edit surgical instead, resynthesizing only the changed span from source audio, original text, and edited text while the surrounding speech is left untouched. The same model performs zero-shot voice cloning from a short prompt recording or a precomputed codebook, plus paralinguistic control through inline tags such as emotion markers, steered by classifier-free guidance scales. Because generation is parallel rather than token by token, first-frame latency lands around 60 ms end to end, which suits live agents and streaming pipelines. Deployment is a bash-scripted Python environment exposing gRPC services with an HTTP gateway, with weights fetched from Hugging Face, and demonstrations cover Chinese and English. Released in June 2026 under Apache 2.0 with no usage restrictions, it occupies a niche little open TTS touches: post-production word-level audio editing combined with low-latency synthesis.