EchoChronos 🥰

Description

EchoChronos (Time Echo Chronos) is a multi-modal style conversational AI assistant based on a large language model, designed to provide users with a new style of conversational experience. The AI integrates RAG, TTS, and other technologies to enable real-time interaction with users, allowing them to immerse themselves in the charm of classic dialogues and dialogues with history. 🥸

Software Architecture

EchoChronos
├─ ChatStyle  # Conversation style module
├─ managers   # Used to provide interfaces for various modules
│  ├─ __init__.py
│  ├─ connect.py  # Connection manager, currently only supports WebSocket
│  ├─ constants.py  # Constants
│  ├─ model.py  # Style dialogue model manager
│  ├─ rag.py  # RAG model manager
│  ├─ runner.py  # Runner manager, used for writing inference logic
│  └─ tts.py  # TTS model manager
├─ RAG  # RAG module
├─ TTS  # TTS module
├─ utils  # Toolkit
├─ inference_torch.py  # Inference code using PyTorch
├─ inference.py  # Inference code using MindSpore
├─ launch.py  # Project entry point
├─ README.en.md
└─ README.md

Installation

python>=3.10

Install ffmpeg

conda install ffmpeg

Install MindNLP

pip install mindnlp==0.4.0

Install 其它依赖

git clone --recursive https://gitee.com/xujunda2024/echochronos.git
cd echochronos
pip install -r requirements.txt

Instructions

1. Set up a Python environment ☝️🤓

2. After successfully installing the dependencies, prepare the configuration file in the format of examples/infer_qwen2_lora_fp32.yaml (be sure to modify the parameters in the configuration file according to your needs).

3. Start the GPT-SOVITS service.

   • Prepare the model: For details, please refer to the README.md file of the GPT-SOVITS project.

   • Start the service:

     cd TTS/GPT-SoVITS-main/GPT_SOVITS
     python Server.py

4. Currently, this project provides three modes of operation, which can be changed by modifying the "isTerminal", "isWebsocket", and "isWebUI" parameters in the YAML file (Replace <your_yaml_path> with your YAML-formatted configuration file).

   • Terminal: python launch.py <your_yaml_path>

   • WebSocket: python launch.py <your_yaml_path>

   • WebUI (recommended 🤩): streamlit run launch.py <your_yaml_path>

MindSpore Users 🤯

GPU Users

• It is recommended to use CUDA 11.6 and cuDNN.

• If you encounter [ERROR] libcuda.so (needed by mindspore-gpu) is not found., please export CUDA_HOME=/path/to/your/cuda

Video Memory

• Due to the accuracy issues of the Qwen2 models in the MindNLP, inference can only be performed using float32, with memory consumption around 46G.

Instructions for MindSpore Users

• MindSpore has two inference script entry points, and the startup methods are as follows:

  • python inference_ms.py --isTerminal or python inference_ms.py --isWebsocket

  • streamlit run webui.py (recommended 🤩)