VibeVoice – AI Text-to-Speech for Real Conversations

Microsoft VibeVoice is an open-source AI text-to-speech system (1.5B parameters) that transforms written text into expressive, multi-speaker audio. It uses a next-token diffusion framework operating at an ultra-low 7.5 Hz frame rate, which lets it understand the full context of a sentence before speaking — resulting in natural rhythm, emotion, and timing. It was accepted as an oral presentation at ICLR 2026.

Most TTS tools process text sentence-by-sentence and produce robotic, monotone output. VibeVoice processes entire passages holistically at 7.5 Hz, enabling it to generate up to 90 minutes of continuous multi-speaker audio with emotionally expressive delivery, natural turn-taking, and realistic breathing pauses. Competing tools typically cap out at a few minutes of mono-speaker output.

VibeVoice supports up to 4 distinct speakers per generation. Each speaker can have a different voice, accent, and emotional style. The AI automatically handles natural turn-taking and overlapping reactions, making the output sound like a real conversation rather than alternating monologues.

Most scripts generate in under 30 seconds. For longer content (30–90 minutes), generation typically completes in 60–90 seconds depending on script complexity and server load. VibeVoice-Realtime, the streaming variant, achieves a first-audible-chunk latency of around 300 milliseconds.

Yes — podcasting is one of VibeVoice's primary use cases. You can paste a two-person or four-person script, assign voices, and get a fully produced podcast-style episode with natural pacing, emotional delivery, and optional background music. Many users create entire podcast series without recording equipment.

VibeVoice-TTS natively supports English and Chinese, with strong cross-lingual voice cloning (e.g., making an English voice speak Chinese). The VibeVoice-ASR transcription model supports 50+ languages. We are actively adding Japanese and Spanish speaker presets to the platform.

Yes. Upload a 5–10 second audio clip of any voice, and VibeVoice will use it as the speaker identity for generation. This works cross-lingually — you can clone an English voice and have it speak Chinese or Japanese. Custom voice uploads are supported directly in the composer.

VibeVoice-TTS supports up to 90 minutes of continuous speech in a single generation. This is far beyond most competitors which cap at 2–5 minutes. For even longer projects, you can chain multiple generations together in your audio editor.

VibeVoice is used by podcasters, audiobook narrators, e-learning course creators, corporate training teams, game developers needing character voices, content marketers producing audio ads, and language learners who need native-sounding listening material. Any workflow that converts written content to audio benefits from VibeVoice.

Yes. All audio generated through VibeVoice AI is yours to use, including for commercial projects — podcasts, marketing, games, education, and more. We do not claim any rights over your output. Please review our Terms of Service for full details.