EmpathyEar: An Open-source Avatar Multimodal Empathetic Chatbot

Hao Fei, Han Zhang, Bin Wang, Lizi Liao, Qian Liu, and Erik Cambria.

This is the repository that contains the source code for the ACL 2024 Demo of EmpathyEar: An Open-source Avatar Multimodal Empathetic Chatbot. EmpathyEar is a pioneering open-source, avatar-based multimodal empathetic chatbot, to fill the gap in traditional text-only empathetic response generation (ERG) systems.

Demonstration

https://github.com/scofield7419/EmpathyEar/assets/18722770/082b238f-dbf2-42a4-a238-abae88fdb110

Workflow

Framework Architecture

Quick Start

Prepare pre-trained models

1. Download ChatGLM3 chekpoints from https://huggingface.co/THUDM/chatglm3-6b/tree/main and place it in the ChatGLM-6B folder.
2. Download pre-trained ChatGLM3 lora checkpoints from https://pan.baidu.com/s/14zzdxyRZL3dqBmI2hJPlIw?pwd=qj4w and place it in the ChatGLM-6B folder. You can also fine-tune ChatGLM by following these steps:

• cd chatglm
• ./scripts/finetune_lora.sh

3. Download the pre-trained StyleTTS2 model on LibriTTS at https://huggingface.co/yl4579/StyleTTS2-LibriTTS/tree/main and place it in the StyleTTS2-LibriTTS folder.
4. Download the pretrained models for EAT and place them in the ckpt,[Utils] folder respectively according to the following steps.

   gdown --id 1KK15n2fOdfLECWN5wvX54mVyDt18IZCo && unzip -q ckpt.zip -d ckpt
   gdown --id 1HGVzckXh-vYGZEUUKMntY1muIbkbnRcd && unzip -q Utils.zip -d Utils

Create conda environment

• conda env create -f environment.yml
• conda activate empathyear

Test on Empathetic Dialogue dataset

• python inference.py
  The generated TTS wav files will be saved in TTS_audio, and the generated talking face videos will be saved in MP4_video.

Acknowledge

We acknowledge these works for their public codes: ChatGLM3, ImageBind, StyleTTS2, EAT.