⏳ Pending community GPU grant

VibeVoice-1.5B
Text-to-Speech Demo

An interactive demo for microsoft/VibeVoice-1.5B — a frontier open-source long-form, multi-speaker text-to-speech model. Enter a script, assign a reference voice to each of up to four speakers, and generate expressive conversational audio.

Status

This is currently a Static holding Space. The application (app.py) is already written for ZeroGPU and will start producing audio as soon as GPU hardware is attached. A community-grant request has been posted in the Community tab.

What it does

Up to 4 distinct speakers, each with a chosen reference voice.
Long-form generation — the 64K-context model can synthesize up to ~90 minutes in one pass.
Runs the official Transformers conversion vibevoice/VibeVoice-1.5B-hf via AutoProcessor / AutoModelForTextToWaveform.

Script format

Speaker 1: Hello everyone, and welcome to the demo.
Speaker 2: Thanks for having me. I'm excited to be here.
Speaker 1: Let's get started.

One turn per line as Speaker N: text (N = 1–4). Lines without a tag default to Speaker 1.

⚠️ Responsible use. VibeVoice is released for research purposes only. Do not use it for voice impersonation without consent, disinformation, or any unlawful purpose. Generated audio is AI-generated — always disclose it. Supported languages: English and Chinese.

VibeVoice-1.5BText-to-Speech Demo

Status

What it does

Script format

VibeVoice-1.5B
Text-to-Speech Demo