This repository contains an ONNX-exported version of the microsoft/VibeVoice-Realtime-0.5B model. This export was manually created to allow cross-platform inference in environments like ONNX Runtime Web (JavaScript) and Flutter (Dart).

All credit for the original model architecture, training, and base weights goes to the Microsoft VibeVoice Team. Please see their original repository for full details and research:

The original weights and software are licensed under the MIT License.

Due to the streaming nature of VibeVoice, the ONNX export is modularized into the following specific components (with accompanying .data files for external weights):

These models are optimized for ONNX Runtime. They can be loaded directly into client-side applications instead of maintaining heavy PyTorch backends. Check the corresponding JS and Flutter demo applications for integration guidance!