VibeVoice

Overview

VibeVoice is Microsoft's open-source family of voice models (MIT licence). As of October 2026 the repository ships:

• VibeVoice-ASR (7B): transcribes up to 60 minutes of audio in one pass and returns who spoke, when, and what. Over 50 languages, no language flag needed, custom hotwords supported.
• VibeVoice-ASR-Streaming (7B): emits text chunk by chunk while audio is still arriving, with hotwords and 10 languages (released September 2026).
• VibeVoice-ASR-BitNet: a quantized CPU build (about 1.6 GB, run through the separate microsoft/VibeASR.cpp engine), no GPU needed.
• VibeVoice-Realtime-0.5B: single-speaker streaming TTS, roughly 200-300 ms to first audio, about 10 minutes of speech per generation. English is the supported language; nine other languages are experimental.
• VibeVoice-TTS-1.5B: long-form multi-speaker TTS (up to 90 minutes, 4 speakers). The weights are on Hugging Face, but Microsoft removed the TTS inference code from the repository in September 2025 after misuse, so the repo cannot run it today.

Important: there is no VibeVoiceTTS or VibeVoiceASR Python class and no pip install vibevoice release from Microsoft. The PyPI package of that name is not the Microsoft project. Install from the GitHub clone, or use the Hugging Face Transformers integration for ASR.

Instructions

Speech recognition with Transformers (simplest path)

VibeVoice-ASR is in Transformers 5.3 and later. The repo itself pins transformers<5, so keep the two in separate virtual environments.