TLDR: Microsoft has launched VibeVoice, an open-source text-to-speech (TTS) AI model designed to create expressive, long-form, multi-speaker conversational audio, such as podcasts, from text. It supports up to 90 minutes of speech with four distinct speakers and utilizes a novel diffusion framework with a Large Language Model (Qwen 2.5) for enhanced scalability and natural turn-taking. The model is available in 1.5 billion and 7 billion parameter variants under an MIT license, intended for research and development.
Microsoft has officially unveiled VibeVoice, a groundbreaking open-source text-to-speech (TTS) AI model that promises to revolutionize the creation of long-form, multi-speaker conversational audio. Announced on August 26, 2025, VibeVoice addresses significant limitations in traditional TTS systems, particularly concerning scalability, speaker consistency, and natural turn-taking in extended audio formats like podcasts.
One of VibeVoice’s most notable capabilities is its ability to synthesize speech for up to 90 minutes, featuring as many as four distinct speakers. This significantly surpasses the typical one or two-speaker limits of many existing models, making it ideal for complex dialogue-rich content.
The core innovation behind VibeVoice lies in its use of continuous speech tokenizers—both Acoustic and Semantic—which operate at an ultra-low frame rate of 7.5 Hz. This design choice efficiently preserves audio fidelity while dramatically boosting computational efficiency, especially crucial for processing lengthy audio sequences.
VibeVoice employs a sophisticated next-token diffusion framework. It leverages a Large Language Model (LLM), specifically the Qwen 2.5 1.5 billion parameter model, to comprehend textual context and the intricate flow of dialogue. A diffusion head then generates high-fidelity acoustic details, ensuring natural and expressive speech.
The model is released in two primary variants: a 1.5 billion parameter model capable of generating up to 90 minutes of audio with a 64k token context length, and a larger 7 billion parameter model with a 32k context length, supporting up to 5 minutes of generation. A 5 billion parameter streaming model is also anticipated for future release.
Currently, VibeVoice supports two languages: English and Chinese Mandarin. Microsoft has made the model available under an MIT license, permitting its use for commercial purposes.
Microsoft emphasizes that VibeVoice is primarily intended for research and development. The company has outlined several out-of-scope uses, including generating voice impressions without explicit recorded consent, creating disinformation or impersonation, producing audio presented as genuine recordings of real people, and real-time or low-latency voice conversations (e.g., for telephone or video conferencing). The model also does not explicitly handle overlapping speech, background noise, music, or other sound effects.
Also Read:
- VIBE VOICE: Advancing Long-Form, Multi-Speaker Speech Synthesis
- Microsoft Outlines Vision for AI-Centric, Multimodal Windows by 2030
Developers and researchers can explore VibeVoice through its GitHub repository and demo spaces, with options for local deployment.
