This is Pocket. You say “Pocket” and a question, it answers out loud, and a face on the OLED listens, thinks and talks. Deepgram does speech to text and the voice, and Claude Haiku writes the answers. It runs on an 18650 and connects over home Wi-Fi or my phone’s hotspot.
Why the XIAO ESP32S3 Sense
My first version ran on a classic ESP32 dev board. With no PSRAM I could only record about 6 seconds at 8 kHz, which wasn’t enough. The S3 Sense has 8 MB of PSRAM for 16 kHz audio buffers, TLS and JSON. It also has a built-in PDM mic and camera, a LiPo charger on the USB-C and it’s tiny, so everything fits in a 8x5.5 cm box.
Libraries
Adafruit SSD1306 + GFX for the face. They’re simple and well documented, and plenty for 128x64.
ArduinoJson 7 to read the Deepgram and Claude replies. Deepgram’s replies are deeply nested, so I had to raise the nesting limit.
IRremote (Armin Joachimsmeyer) to learn and replay remote buttons.…