Developer Builds Voice-Activated 10.3-Inch E-Paper Memo Board

Mason and Mengdu published a maker project on Hackster.io demonstrating a voice-activated reminder board built on a Seeed Studio reTerminal E1003, an ESP32S3-based development board with a 10.3-inch e-paper display running at 1872 by 1404 pixels in 16-level grayscale. Pressing the onboard KEY0 button starts audio capture through the board's built-in PDM microphone; releasing it uploads the recording to a cloud API for transcription and reminder extraction. The completed memo appears as a formatted card on the e-paper screen in five to eight seconds.

The project uses Groq's free API tier throughout, keeping operating costs at zero within standard rate limits. Whisper Large V3 Turbo handles speech recognition from 16 kHz, 16-bit mono PDM audio. The transcript goes to LLaMA 3.3 70B Versatile with a system prompt that includes the current date, time, and weekday, so the model can resolve natural-language deadlines such as "by Thursday afternoon" or "next Monday" into a structured JSON object with the memo text, exact due time, and a human-readable label such as "Tomorrow morning." Temperature is set to 0 for deterministic parsing. The display shows eight cards simultaneously; a GT911 capacitive touchscreen lets users tap a checkbox on completed items, which turn gray and sort to the bottom, with the oldest completed entry replaced when a ninth memo arrives.

For testing without API calls, the project includes a Python gateway server that runs on a laptop in mock mode with no external dependencies, or with real transcription using faster-whisper and a local LLM via Ollama. The same firmware supports two smaller reTerminal variants (E1001 and E1002, both 800x480) that lack a touchscreen; on those, the display holds four cards and automatically rotates out the oldest entry by due time. Source code and firmware are published under MIT license at GitHub (limengdu/ePaper-Voice-Memo).

The authors document a non-obvious hardware constraint: e-paper refreshes via the IT8951 controller take hundreds of milliseconds and overflow the I2S DMA ring buffer (capacity approximately 256 ms of audio), corrupting the start of any recording that follows a screen update. The fix is to defer all display changes until after audio is safely copied to PSRAM, using the onboard LED and buzzer for recording feedback instead of a visual status update. This pattern - delaying peripheral I/O that shares CPU cycles with time-sensitive data capture - applies broadly to constrained embedded AI systems.