Mini-Omni2

🤗 Hugging Face | 📖 Github | 📑 Technical report

If you like little Omni2, please give us a star⭐ and cite our paper!

Introduction

Mini-Omni2 is an omni-interactive model. It can understand image, audio and text inputs and has end-to-end voice conversations with users. Featuring real-time voice output, omni-capable multimodal understanding and flexible interaction ability with interruption mechanism while speaking.

Updates

2024.10: Release the model, technical report, inference and chat demo code.

Features

✅ Multimodal interaction: with the ability to understand images, speech and text, just like GPT-4o.

✅ Real-time speech-to-speech conversational capabilities. No extra ASR or TTS models required, just like Mini-Omni.

Demo

NOTE: need to unmute first.

mini-omni2-demo.mp4

ToDo

update interruption mechanism
visual-assistant model and data

Install

Create a new conda environment and install the required packages:

conda create -n omni python=3.10
conda activate omni

git clone https://github.com/gpt-omni/mini-omni2.git
cd mini-omni2
pip install -r requirements.txt

Quick start

Interactive demo

start server

NOTE: you need to start the server before running the streamlit or gradio demo with API_URL set to the server address.

sudo apt-get install ffmpeg
conda activate omni
cd mini-omni2
python3 server.py --ip '0.0.0.0' --port 60808

run streamlit demo

NOTE: you need to run streamlit locally with PyAudio installed.

pip install PyAudio==0.2.14
API_URL=http://0.0.0.0:60808/chat streamlit run webui/omni_streamlit.py

Local test

conda activate omni
cd mini-omni2
# test run the preset audio samples and questions
python inference_vision.py

Mini-Omni2 Overview

1. Multimodal Modeling: We use multiple sequences as the input and output of the model. In the input part, we will concatenate image, audio and text features to perform a series of comprehensive tasks, as shown in the following figures. In the output part, we use text-guided delayed parallel output to generate real-time speech responses.

2. Multi-stage Training: We propose an efficient alignment training method and conduct encoder adaptation, modal alignment, and multimodal fine-tuning respectively in the three-stage training.

FAQ

1. Does the model support other languages?

No, the model is only trained on English. However, as we use whisper as the audio encoder, the model can understand other languages which is supported by whisper (like chinese), but the output is only in English.

2. Error: can not run streamlit in local browser, with remote streamlit server

You need start streamlit locally with PyAudio installed.

Acknowledgements

Qwen2 as the LLM backbone.
litGPT for training and inference.
whisper for audio encoding.
clip for image encoding.
snac for audio decoding.
CosyVoice for generating synthetic speech.
OpenOrca and MOSS for alignment.

Citation

@article{xie2024miniomni2opensourcegpt4ovision,
      title={Mini-Omni2: Towards Open-source GPT-4o with Vision, Speech and Duplex Capabilities}, 
      author={Zhifei Xie and Changqiao Wu},
      year={2024},
      eprint={2410.11190},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      journal={ArXiv},
      volume={abs/2410.11190},
}

DuStaRing/mini-omni2