VibeVoice-ASR-Streaming-AWQ-4bit

VibeVoice-ASR-Streaming-7B

AWQ 4-bit quantization

This repository is a Gaia-produced 4-bit AWQ quantization of microsoft/VibeVoice-ASR-Streaming-7B. The Qwen2 language-model decoder is quantized with AutoAWQ using 4-bit, group-size 128, GEMM-compatible weights. The VibeVoice acoustic and semantic tokenizers/connectors remain BF16 to preserve multimodal streaming ASR behavior. The original model is a BF16 checkpoint reported as ~9B parameters despite the repository's 7B name.

The model requires the VibeVoice vLLM plugin from Microsoft's VibeVoice repository and a vLLM build with AWQ support. The quantized decoder is in decoder-awq/; the root configuration preserves the streaming architecture and points the plugin to it.

Quantization used 128 ASR/transcription-oriented calibration prompts. This is weight-only quantization; speech-tokenizer weights were not quantized.

VibeVoice-ASR-Streaming is a unified streaming ASR model that transcribes Who (Speaker) said What (Content), with support for Customized Hotwords and 10 languages.

➡️ Code: microsoft/VibeVoice<br> ➡️ Demo: VibeVoice-ASR-Streaming<br>

<p align="left"> <img src="figures/VibeVoice_ASR_Streaming_architecture.png" alt="VibeVoice-ASR-Streaming Architecture" height="250px"> </p>

🔥 Key Features

• 📝 Streaming Speaker-Attributed Transcription: Continuously transcribes who said what as speech arrives.

• 👤 Customized Hotwords: Users can provide customized hotwords, such as names and technical terms, to improve recognition of domain-specific content.

• 🌍 Multilingual Support: It supports Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.

Technical Report

📄 VibeVoice-ASR-Streaming Technical Report

Evaluation

<p align="center"> <img src="figures/VibeVoice_ASR_Streaming_results.png" alt="VibeVoice-ASR-Streaming Results" width="80%"> </p>

Installation and Usage

Please refer to the GitHub repository.

License

This project is licensed under the MIT License.

Contact

This project was conducted by members of Microsoft Research. We welcome feedback and collaboration from our audience. If you have suggestions, questions, or observe unexpected/offensive behavior in our technology, please contact us at [email protected]. If the team receives reports of undesired behavior or identifies issues independently, we will update this repository with appropriate mitigations.