Build a Raspberry Pi 5 Offline Voice Assistant

Flash Raspberry Pi OS with SSH enabled

Use Raspberry Pi Imager to write current 64-bit Raspberry Pi OS to the microSD card. Set a hostname, username, password, Wi-Fi and SSH in Imager before first boot.

• Use 64-bit Raspberry Pi OS.
• Keep networking enabled until every package, model and voice file is cached.
• Record the hostname and login before ejecting the card.

Raspberry Pi Imager →

Dry-fit the cooler, battery, riser and Whisplay HAT

Install active cooling on the Pi 5, align the PiSugar 3 Plus underneath, add the 2×20 stackable header and seat the Whisplay HAT above the cooler. Check every pin before applying pressure.

• Disconnect all power before assembly.
• Never offset the 40-pin connector by one row.
• Do not bend a cooler fastener unless you understand the mechanical risk shown in the creator video.
• Confirm the HAT sits level and no metal part contacts the battery board.

Install and test the current Whisplay driver

SSH into the Pi, clone the official driver, pin the revision reviewed by LocalClaw, install it and reboot. Run the hardware demo before adding any AI software.

git clone https://github.com/PiSugar/Whisplay.git --depth 1
cd Whisplay
git checkout a695240f866257033414697d47c3aa4c5f9b749c
sudo bash install_driver.sh
sudo reboot

Note: After reboot, return to ~/Whisplay/example, install its requirements and run bash run_test.sh. Do not continue until the display, button, LEDs, microphone and speaker pass the hardware test.

Install the PiSugar power manager

Install the official release-channel power manager, select the PiSugar 3 model when prompted and confirm that the battery service sees the board.

cd ~
wget -O pisugar-power-manager.sh https://cdn.pisugar.com/release/pisugar-power-manager.sh
bash pisugar-power-manager.sh -c release
systemctl status pisugar-server --no-pager

Note: The official manager exposes its local web interface on port 8421. Change its default credentials before exposing that port beyond your trusted LAN.

Clone and build the reviewed Whisplay chatbot

Clone the application, pin the current reviewed revision, install its system and Python dependencies, then load the Node environment added to your shell.

cd ~
git clone https://github.com/PiSugar/whisplay-ai-chatbot.git
cd whisplay-ai-chatbot
git checkout ee7301b2f7693111ee016f81c48a75d3387a7326
bash install_dependencies.sh
source ~/.bashrc
bash build.sh

Note: The current repository has evolved since the 2025 video. Pinning the reviewed commit keeps the commands and configuration fields on this page reproducible.

Install and cache the three offline engines

Install Ollama and pull Qwen3 1.7B, install Whisper and cache its tiny English model, then install Piper HTTP and download one English voice.

cd ~/whisplay-ai-chatbot
bash scripts/install_ollama.sh
ollama pull qwen3:1.7b
python3 -m pip install -U openai-whisper --break-system-packages
python3 -c "import whisper; whisper.load_model('tiny')"
python3 -m pip install 'piper-tts[http]' --break-system-packages
mkdir -p /home/pi/piper
cd /home/pi/piper
python3 -m piper.download_voices en_US-amy-medium

Note: The /home/pi paths match the official defaults and the creator setup. If your Raspberry Pi username is not pi, use your actual home directory consistently in .env.

Configure the local voice pipeline

Run the official configuration wizard, choose Whisper for ASR, Ollama for the LLM and Piper HTTP for TTS, then verify the critical values in .env.

• ASR_SERVER=whisper
• LLM_SERVER=ollama
• TTS_SERVER=piper-http
• OLLAMA_MODEL=qwen3:1.7b
• SERVE_OLLAMA=true
• WHISPER_MODEL_SIZE_OR_PATH=/home/pi/.cache/whisper/tiny.pt
• PIPER_HTTP_MODEL=/home/pi/piper/en_US-amy-medium
• ENABLE_THINKING=false for the first full test

cd ~/whisplay-ai-chatbot
whisplay configure
grep -E '^(ASR_SERVER|LLM_SERVER|TTS_SERVER|OLLAMA_MODEL|SERVE_OLLAMA|WHISPER_MODEL_SIZE_OR_PATH|PIPER_HTTP_MODEL|ENABLE_THINKING)=' .env

Note: If the wizard does not expose one of these advanced fields, edit .env directly. Never add a cloud API key for this fully local configuration.

Run the chatbot and test the complete audio loop

Start the application in the foreground, press the Whisplay button, speak a short request and wait for the display and speaker response. Keep this foreground test visible until every stage works.

cd ~/whisplay-ai-chatbot
bash run_chatbot.sh

Note: The first Whisper and Ollama request is slower because processes and model data are warming up. Thinking mode should remain disabled during this baseline test.

Disconnect networking and enable startup only after proof

Stop the foreground process, disconnect Wi-Fi or Ethernet, repeat the same voice request, then install the startup service once the offline test succeeds.

• Qwen3, Whisper and the Piper voice must already be present locally.
• A successful disconnected request proves the demonstrated pipeline, not every optional plugin.
• startup.sh switches the Pi to multi-user mode and disables the graphical desktop.
• If whisplay-daemon is installed, follow the current repository warning instead of creating a competing legacy chatbot service.

cd ~/whisplay-ai-chatbot
bash startup.sh
systemctl status chatbot --no-pager
journalctl -u chatbot -n 100 --no-pager