VibeVoice

Microsoft Research uses this repository to host a family of speech models rather than a single TTS system. The current lineup centers on VibeVoice-ASR, a 7B recognizer that takes up to an hour of audio in one pass and returns text labeled with speaker and timing, accepts hotword lists for names and jargon, and covers more than 50 languages; streaming and CPU-only BitNet variants exist, the latter through the separate VibeASR.cpp engine. On the synthesis side, VibeVoice-Realtime-0.5B streams mostly English speech from incoming text with first audio in roughly 300 ms. The original multi-speaker model for podcast-length synthesis still has weights on Hugging Face, but Microsoft pulled its inference code from the repo in September 2025 after misuse. Code is MIT licensed; the documented setup is a pip install inside an NVIDIA PyTorch container, ASR also loads through Hugging Face Transformers or a vLLM plugin, and fine-tuning scripts are included for ASR. Microsoft states the models are meant for research and advises against commercial deployment without further testing. Suits researchers prototyping transcription or voice agents.