What’s in the box:

• ESP32-S3 control board (16MB Flash, 8MB PSRAM)
• 2.0” TFT-SPI ST7789 display module
• I2S digital microphone module (INMP441)
• I2S amplifier module (MAX98357A) with 3W speaker
• USB-C cable for flashing and power
• Standoffs and screws for modular assembly

[DEPLOYMENT_SEQUENCE]

Step 1: Assemble the Hardware (5 minutes)

The kit is modular. Follow the pin markings — each module only connects one way.

ESP32-S3 Board
├── Display  → SPI header (labeled "TFT")
├── Mic      → I2S header (labeled "I2S IN")
├── Speaker  → I2S header (labeled "I2S OUT")
└── USB-C    → Power + Programming

Snap together. No soldering. Done.

Step 2: Flash the Firmware

We’ll use XiaoZhi-ESP32 — the open-source firmware with 26,000+ GitHub stars that powers this exact kit. It supports both DeepSeek and OpenAI APIs out of the box, plus custom WebSocket endpoints for local LLMs.

2.1 Install esptool

pip install esptool

2.2 Download the firmware

git clone https://github.com/78/xiaozhi-esp32.git
cd xiaozhi-esp32

2.3 Flash to the ESP32-S3

Connect the USB-C cable. On Windows (from WSL):

esptool.py --chip esp32s3 --port /dev/ttyUSB0 erase_flash
esptool.py --chip esp32s3 --port /dev/ttyUSB0 write_flash -z 0x0 build/bootloader/bootloader.bin 0x10000 build/partition_table/partition-table.bin 0x20000 build/xiaozhi.bin

On Windows directly (PowerShell):

esptool.py --chip esp32s3 --port COM3 erase_flash
esptool.py --chip esp32s3 --port COM3 write_flash -z 0x0 .\build\bootloader\bootloader.bin 0x10000 .\build\partition_table\partition-table.bin 0x20000 .\build\xiaozhi.bin

Step 3: Configure Wi-Fi + Tailscale

After flashing, the ESP32-S3 creates a hotspot. Connect to it from your phone or laptop:

1. Join Wi-Fi network: XiaoZhi-Config
2. Open http://192.168.4.1 in a browser
3. Enter your home Wi-Fi credentials
4. Enter your LLM server’s Tailscale IP (e.g., 100.xxx.xxx.xxx)
5. Set the WebSocket port (default: 8080)

The ESP32 will reboot and connect to your network.

Step 4: Set Up the Local LLM Server

On your NukBox or any machine with a GPU:

Option A: Ollama (Simplest)

curl -fsSL https://ollama.ai/install.sh | sh
ollama pull llama3:8b
ollama pull qwen2.5:7b

Option B: vLLM (Production)

pip install vllm
vllm serve mistralai/Mistral-7B-Instruct-v0.3 --host 0.0.0.0 --port 8000

Step 5: Install Speech-to-Text + Text-to-Speech

On the LLM server, run this bridge:

# STT: Whisper
pip install faster-whisper
# TTS: Piper (offline, fast)
pip install piper-tts

Or use the XiaoClaw firmware which bundles STT/TTS directly on the ESP32:

git clone https://github.com/beancookie/xiaoclaw.git

Step 6: Wire It All Together with n8n

Create an n8n workflow (or a simple Python WebSocket server) that:

ESP32 Voice Input → WebSocket → n8n/Webhook
    → Whisper STT (audio → text)
    → Ollama/vLLM (text → response)
    → Piper TTS (response → audio)
    → WebSocket back to ESP32
    → Speaker plays response

This keeps your voice data air-gapped — it never leaves your home network.

Step 7: Test the Wake Word

Say your wake word. The ESP-SR engine runs entirely on the ESP32-S3 — no internet, no cloud. It fires instantly.

First test: “Hey Dragon, what’s the weather?” Response: “I don’t have internet access, but your RTX 3060 is at 42°C and you have 3 unread emails.”

[ADVANCED_MODS]

Connect to Alexa as a Speaker

The ESP32’s MAX98357A amplifier outputs via I2S. To use your Alexa Echo as the speaker:

1. Option A: ESP32 I2S → 3.5mm aux jack → Alexa line-in port
2. Option B: ESP32 Bluetooth → pair to Alexa as Bluetooth speaker
3. Option C: Use the onboard 3W speaker for a fully self-contained unit

3D Printed Chassis

Design a custom enclosure on your Bambu A1:

• Cutout for the 2” TFT display (visible through the shell)
• Mic port aligned with the INMP441
• Speaker grille for sound passthrough
• Ventilation slots for the ESP32 (enclosed PLA needs airflow)

Add LED Eyes

Wire WS2812B NeoPixels to GPIO 48. Program animated eye patterns that “blink” while the LLM thinks.

<PartsList parts={[ { name: “LAFVIN ESP32-S3 AI Chatbot Kit”, link: “https://amzn.to/lafvin-esp32-s3-kit”, price: “$51.03” }, { name: “N100 Mini PC (LLM Server)”, link: “https://amzn.to/beelink-s12-pro”, price: “$169.00” }, { name: “USB-C Data Cable”, link: “https://amzn.to/usbc-data-cable”, price: “$7.99” }, { name: “WS2812B NeoPixel Ring (optional)”, link: “https://amzn.to/neopixel-ring”, price: “$9.99” }, { name: “3.5mm Aux Cable (Alexa mod)”, link: “https://amzn.to/aux-cable”, price: “$5.99” } ]} />

<RelatedTools tools={[ { title: “XiaoZhi-ESP32”, link: “https://github.com/78/xiaozhi-esp32”, description: “26K+ star open-source firmware for ESP32 voice chat. Pre-loaded MCP support.” }, { title: “XiaoClaw”, link: “https://github.com/beancookie/xiaoclaw”, description: “Local AI Agent firmware with tool calling, memory, and autonomous task execution.” }, { title: “Ollama”, link: “https://ollama.ai”, description: “Run Llama 3, Qwen 2.5, Mistral locally with one command.” }, { title: “Tailscale”, link: “https://tailscale.com”, description: “Mesh VPN connecting your ESP32, server, and phone on one private network.” }, { title: “n8n”, link: “https://n8n.io”, description: “Workflow automation to orchestrate STT → LLM → TTS pipelines.” } ]} />

[TRouBLeSHooTiNg]

[NEXT_STEPS]

Now that your voice assistant is live, extend it:

1. [Add tool calling] — Let it control smart home devices, check system stats, or trigger n8n workflows
2. [3D print a dragon chassis] — See our Dragon Smart Speaker build
3. [Run fully offline] — Switch to XiaoClaw firmware for local LLM inference directly on ESP32-S3
4. [Build a cyberdeck terminal] — See our NerdDeck build guide

Transmission complete. Your voice assistant is no longer a corporate listening device. It’s yours.