A wearable AI that listens to your entire life

What it does

Owl is a full-stack system for wearable, always-on AI. It ingests continuous audio and location data from low-power devices—ESP32 boards, Sony Spresense, Apple Watch, or the project’s own Bee puck—and forwards it to a server that transcribes conversations, detects speech endpoints, and generates summaries. You interact with the results through an iOS or web app, effectively giving you a searchable, AI-curated memory of your day.

The interesting bit

The architecture cleverly uses your iPhone as a network relay: tiny BLE wearables stream AAC audio to the iOS app, which forwards packets over a socket to the server for real-time processing. This keeps the wearable power-efficient and small while the phone handles the heavier internet lift. It is essentially an open-source life-logging pipeline dressed up as a personal assistant.

Key highlights

• LLM inference can run offline via Ollama, or you can fall back to commercial APIs like GPT-4.
• Supports a wide range of capture hardware, from off-the-shelf Apple Watches to semi-custom ESP32 and Sony Spresense boards.
• Includes speaker verification so the system can distinguish the wearer’s voice from background conversation.
• The server uses voice-activity detection to chop continuous audio streams into discrete conversation segments for summarization.
• The README is admirably upfront about privacy risks, warning that default HTTP setups leak audio and transcripts in plaintext and urging users toward HTTPS or reverse proxies.

Caveats

• Despite the “run completely locally” banner, the technical walkthrough notes that real-time streaming transcription currently relies on Deepgram, with a local alternative still planned.
• Only audio and location capture are implemented today; image and video support are explicitly flagged as coming very soon.
• Android client support is also listed as coming soon, so mobile capture and interaction are currently limited to iOS and the web.

Verdict

Developers and HCI researchers interested in building a personal, always-on memory prosthetic will find a rare fully open stack to hack on. If you want a polished, consumer-grade product with turnkey privacy, this is still an early experiment with visible rough edges.