VibeVoice

VibeVoice is Microsoft’s open-source frontier voice AI family with 54,900+ stars on GitHub, built to turn text scripts and conversations into natural, long-form spoken audio.

Overview

Most speech models handle one speaker and a few seconds of audio. VibeVoice targets the opposite end: expressive, podcast-style recordings up to 90 minutes long with as many as four distinct speakers whose voices stay consistent from start to finish. It generates speech from continuous speech tokenizers operating at an ultra-low 7.5 Hz frame rate, which keeps long sequences efficient, and pairs them with a next-token diffusion head that adds the fine acoustic detail that makes output sound performed rather than read aloud. The family also includes a compact realtime model designed for low-latency streaming text-to-speech. Released by Microsoft Research under the MIT license, VibeVoice supports English and Chinese with cross-lingual voice transfer, and it remains actively developed — the repository was updated in October 2026.

Key Features

• Long-form multi-speaker TTS – synthesize up to 90 minutes of audio with up to 4 speakers in one pass.
• Conversational, expressive output – designed for podcasts, dialogues and audiobooks rather than flat narration.
• Realtime streaming model – a smaller variant built for low-latency, interactive text-to-speech.
• Efficient tokenizers – continuous speech tokenizers at 7.5 Hz keep long-form generation fast and coherent.
• Bilingual and cross-lingual – English and Chinese voices, including speaking one language with another’s voice.
• Fully open – code and model weights are public on GitHub and Hugging Face under the MIT license.
• Runs locally – generate on your own GPU with PyTorch, with no per-minute API fees.

💳 Pricing

VibeVoice is free and open source under the MIT license. You download the models and run them on your own hardware, so the only cost is the GPU time you supply; Microsoft offers it as a research release for the community to build on.

Who Is It For

Podcasters, audiobook producers, voice-app developers and speech researchers who need long, multi-speaker, human-sounding audio — and who want to generate it locally instead of paying a cloud TTS service by the minute.