A simple voice chatbot built on ASR -> LLM -> TTS. Press Enter to record your speech, the audio is transcribed by PaddleSpeech, answered by a local LLM served through Ollama (qwen2.5:7b), and the reply is synthesized and played back.
A GPU is required — the ASR is hard-coded to run on device='gpu', and the PaddlePaddle build installed by install.sh targets CUDA 12.0.
- NVIDIA GPU with a driver supporting CUDA 12.0
sox(provides thereccommand used for recording)alsa-utils(provides theaplaycommand used for playback)
sudo apt update
sudo apt install -y sox alsa-utilsPython 3.10 is recommended:
conda create -n llm python=3.10
conda install -n llm -c conda-forge cudnn=8.9.7.29
conda activate llmsh install.shThis installs PaddleSpeech (ASR + TTS), LangChain and the Ollama integration.
# install Ollama first
ollama pull qwen2.5:7bRun the following script to chat with LLM:
conda activate llm
python voice_assistant.py