VibeVoice's feature set is designed around one problem: generating long, multi-speaker audio that sounds like a real conversation. Most TTS models max out at single-speaker clips under a minute. VibeVoice generates 90-minute dialogues with 4 speakers. Every architectural decision targets this capability.

90-Minute Long-Form Generation

Most TTS models degrade after 30-60 seconds. VibeVoice generates up to 90 minutes of continuous speech without quality loss. The ultra-low 7.5 Hz tokenizer is the key: it produces fewer tokens per second of audio, keeping long sequences within the model's context window. A 90-minute podcast episode that would require thousands of separate clips with other models generates in one pass with VibeVoice.

4-Speaker Consistency

Four distinct speakers maintain consistent voice identities across the entire output. Speaker 1 sounds the same on minute 1 and minute 90. Turn-taking happens naturally: speakers respond to each other with appropriate timing, pauses, and prosody shifts. This is not separate clips stitched together. The model generates the full multi-speaker conversation as one continuous output.

7.5 Hz Ultra-Low Frame Rate Tokenizer

The core innovation. Most TTS systems tokenize audio at 50-100 Hz. VibeVoice uses continuous speech tokenizers (Acoustic and Semantic) at 7.5 Hz. This 7-13x reduction in token rate makes it computationally feasible to process sequences long enough for full podcast episodes. The tokenizers preserve audio fidelity despite the lower frame rate through continuous (not discrete) representations.

Next-Token Diffusion

The architecture combines an LLM backbone (for understanding text context and dialogue flow) with a diffusion head (for generating high-fidelity acoustic details). The LLM plans what to say and how to say it. The diffusion head handles the fine-grained audio synthesis. This hybrid approach produces more natural-sounding speech than pure autoregressive or pure diffusion methods.

Voice Prompting

Provide a short audio sample and VibeVoice captures the speaker's vocal characteristics. The cloned voice stays consistent across the full output duration. You can define each of the 4 speakers with a different reference, creating a complete cast of distinct characters for podcast or audiobook production.

MIT License

Fully open source under the MIT License. Model weights, inference code, and training details are available on HuggingFace and GitHub. Full commercial rights. No usage restrictions. You can deploy it, modify it, fine-tune it, and build commercial products with it.