Python Speech-to-Speech Assistant

This repo contains a ready to use speech-to-speech assistant designed to run on edge devices like the Jetson Nano. It uses ollama in combination with vosk and piper to achieve real-time conversational capabilities with minimal system requirements.

Installation

1. Install ollama - https://ollama.com/download

2. Install jq (can be skipped - needed only for running ./install.sh).

   For Debian-based systems:

   sudo apt-get install jq

3. Clone the repository:

   git clone https://github.com/M1chol/atom-assistant
   cd atom-assistant

4. You then need to manually install all the ollama models you want to use:

   ollama serve
   ollama run <MODEL_NAME>

5. Update config.json with your values. You need to change:

   • ollama_model_name to match a model available in your ollama server.
   • LLM system prompt to match your language and requirements (ollama_system_prompt).
   • Microphone sample rate to match your hardware (stt_config/microphone_samplerate).
   • Speech-to-text vosk model to match your language (stt_config/model).
   • Text-to-speech piper model to match your language (tts_config/model).

6. Run the install script:

   ./install.sh

Running

After running the install script, execute:

python main.py

If the script fails to activate the Python environment (could happen in non-desktop environments), do it manually:

source .venv/bin/activate
python main.py

Tweaking Default config.json Values

In config.json, you can change the reading speed. If your LLM is not fast enough and is blocking speech generation, you can increase the tts_config/length_scale to slow down the reading speed.

noise_scale and noise_w_scale control respectivly audio and speaking variation. You can also change the volume of the output speech.

Singular use examples

I have also included example files for educational puposes for the used ollama, piper and vosk libraries. They all in some way expand on simple examples provided by the library contributors.