Reading the world aloud: scene description and OCR on Echo 1

Obstacle alerts keep you safe; scene description and OCR make the world legible. Ask Echo 1 'what's in front of me?' and it answers in a sentence. Point it at a sign, a menu, or a medicine label and ask it to read.

Scene description with a 500M-parameter vision model

SmolVLM2-500M runs locally via llama.cpp on the Pi 5. When the user asks, the vision service requests a fresh frame from the camera over the event bus, runs the model, and speaks a concise description: 'A hallway with a chair on the left and an open door ahead.' Small models keep latency in conversational range — and honest brevity beats florid hallucination.

OCR that reads like a person, not a scanner

Tesseract handles text recognition, but raw OCR output is unreadable aloud — column breaks, stray characters, ALL-CAPS headers. Echo 1 post-processes the text into clean sentences before speaking, and reads in document order so a menu sounds like a menu.

Find my object

A dedicated tool combines detection with direction mapping: ask 'find the chair' and Echo 1 scans the scene and answers with a clock direction — 'Chair at 10 o'clock.' Small feature, disproportionate independence.

• All three capabilities run offline — in a basement, on a plane, anywhere.
• Photos are processed in memory and discarded; nothing is uploaded, nothing is stored.
• Vision requests never block the safety engine, which keeps its own camera pipeline priority.