VibeVoice
Open speech recognition and text-to-speech models from Microsoft
What it is
A family of open-weight voice models covering long-form speech recognition and text-to-speech, built on low frame rate speech tokenizers with an LLM and diffusion head. The ASR models transcribe up to an hour of audio in one pass with speaker labels, timestamps, custom hotwords, and streaming output across many languages, while the realtime TTS model generates speech from streaming text input. Researchers and developers use the weights, fine-tuning code, and vLLM or CPU inference paths to add transcription or speech synthesis to their own projects.