[New Model] Add support for VibeVoice TTS family (Realtime-0.5B and TTS-1.5B)

Model Overview

VibeVoice is Microsoft's open-source frontier voice AI family. This issue tracks support for their two TTS models:

Model	Weights	Size	Key Capability
VibeVoice-Realtime-0.5B	Public	0.5B	Real-time streaming TTS, 7.5Hz token rate
VibeVoice-TTS-1.5B	Public	1.5B	Long-form multi-speaker TTS (up to 90min, 4 speakers)

Architecture

Both models use a next-token diffusion framework:

• An LLM backbone (Qwen2.5-based) understands textual context and generates continuous speech tokens
• A diffusion head generates high-fidelity acoustic details
• Ultra-low frame rate of 7.5 Hz — efficient for long sequences

The streaming model (VibeVoice-Realtime-0.5B) maps well to vllm-omni's async_chunk pipeline and is a natural fit for the existing TTS serving infrastructure.

Existing vLLM Integration

Microsoft already provides a vllm_plugin/ in their repo with:

• model.py — vLLM multimodal registry integration (ASR-focused currently)
• inputs.py — input processing
• Stage configs and test scripts

This gives us a solid starting point for the TTS integration.

Motivation

• Realtime-0.5B: Streaming-first design aligns directly with vllm-omni's async_chunk architecture. Low model size makes it accessible on consumer GPUs.
• TTS-1.5B: Long-form multi-speaker generation (up to 90 minutes) is a capability no current vllm-omni TTS model covers. The model weights remain publicly available on HuggingFace.

References

• GitHub: https://github.com/microsoft/VibeVoice
• HuggingFace collection: https://huggingface.co/collections/microsoft/vibevoice-68a2ef24a875c44be47b034f
• TTS paper: https://arxiv.org/pdf/2508.19205
• Realtime Colab demo: https://colab.research.google.com/github/microsoft/VibeVoice/blob/main/demo/vibevoice_realtime_colab.ipynb

Implementation Notes

• Both models share the same vibevoice package and diffusion head architecture
• Realtime model config: configuration_vibevoice_streaming.py / modeling_vibevoice_streaming.py
• TTS-1.5B model config: configuration_vibevoice.py / modeling_vibevoice.py
• Stage configs will need to be added under vllm_omni/model_executor/stage_configs/

cc @linyueqian