HealthChat

Our project aims to develop a locally-operated multilingual conversational AI that integrates several advanced technologies to enhance mental health support. These include speech-to-text (STT), Chinese large language models (LLM), and text-to-speech (TTS). Additionally, we have incorporated Differentiable Digital Signal Processing - Singing Voice Conversion (DDSP-SVC) technology, enabling the AI to transform voice characteristics and respond in a comforting, familiar voice. Our model, fine�tuned on chinese psychotherapy datasets, can provide personalized counseling services

1. Speech Recording or Text Input: Users can record their speech via the web interface or input text through the message box.
2. Speech-to-Text: Whisper converts the recorded speech to text.
3. Generating System Response: The converted text is input into the Fine-tuned chinese LLM to generate a response.
4. Text-to-Speech: Bark converts the LLM-generated response into human-like speech.
5. Personalized Voice Training and Cloning: Users can train their desired voice using the DDSP-SVC through the interface, and replace the system's voice with it.

Setup

The recommended version of python is 3.8 to avoid gradio error messages.

Finetuned Taiwan-LLM-13B

We use our own mental dataset to finetune model , can download this model in here: https://huggingface.co/huangyt/module_v7

cd taiwan_llama/module
git Ifs install
git clone https://huggingface.co/huangyt/module_v7

Run

python launch.py

Demo