ESP32-S3 voice chatbot built twice, from perfboard to custom PCB

Nov 19, 2026 Last reply: 2 hours ago 3 Replies
Jump to replies (3)

The detail that surprised me most is how expressive this little machine is: the tiny OLED shows a scrolling face icon with dynamic expressions that match what the chatbot is saying, while the board answers spoken questions out loud in clear English. That is a lot of personality from a display smaller than a stamp.



The maker, KOKENSHA TECH, actually built this twice. The first prototype is a perfboard around an ESP32-S3 DevKitC-1 with jumper wires, a 0.91-inch 128x32 I2C OLED, an INMP441 MEMS microphone, a MAX98357A I2S amplifier, tactile buttons and a small external speaker. The second is a dual-board custom PCB with an ESP32-S3-WROOM-1 module, a CH340C USB-to-UART converter, an AMS1117 3.3V regulator, an OLED, the microphone, the amplifier, an SMD speaker and header breakout pins. The KiCad design files are shared on GitHub, and the boards were produced with support from PCBWay.



The software path is refreshingly short. Espressif's esptool gets installed with pip, the precompiled archive v2.2.4_bread-compact-wifi.zip comes from the GitHub releases page, and merged-binary.bin is written at offset 0x0 at 460800 baud, with the board showing up as USB-SERIAL CH340 on COM19. The configuration portal lives at 192.168.4.1, the device joins the network within roughly 10 to 30 seconds, and a six-digit code spoken by the board registers it on xiaozhi.me. The dialogue language was switched from Chinese to English, with the Harmony (en-US) voice selected.



In the demo it fielded questions on quantum physics, noisy circuits and monetizing open source, leaning on cloud models like Qwen or DeepSeek while the ESP32 handles audio in and out. One note: if configuration settings do not take effect right away, the board needs a reset, so a reboot after changes seems part of the routine.



Project by KOKENSHA TECH. Watch it on their YouTube channel:

formatting link


Flashing a precompiled merged binary straight to offset 0x0 strikes me as a smart shortcut. It skips building a toolchain entirely, so a first-time flasher can go from download to talking hardware in one command. The trade-off is that you are locked to the released configuration until you compile your own image, which for a first run seems like a fair deal.

On the custom PCB, did the maker keep the microphone placed away from the amplifier and the SMD speaker? With everything on one dual-board layout, I would expect the mic to sit fairly close to the audio output, and I am curious how the placement was handled there.

Since the documentation shows MCP can drive local peripherals like LEDs, servos and GPIO pins, and even tie into Home Assistant, a natural next step for this exact board would be wiring something to those header breakout pins. Asking the chatbot out loud to toggle a relay or move a servo would show off the control side, not just the conversation side.

Join the Discussion

Have something to add? Share your thoughts — no account required.

Didn't find your answer?

Ask the community — no account required