◉
Listen
Capture voice through a USB microphone and manage interaction on an ESP32-S3.
Prototype-stage embedded AI project · Taiwan
A little companion.A conversation beyond the screen.
Our long-term goal is a small, expressive AI companion you can talk to in the physical world. Starting with an ESP32-S3 conversation-device prototype, we plan to connect Claude-powered dialogue with a voice, an animated face and responsive hardware.
ESP32-S3Embedded controller
Voice-firstMicrophone → response
Modular I/OAudio · display · actions
01 / Product
A voice, a face, and a sense of continuity.
The vision is a characterful physical companion: one that can hold a conversation, express its interaction state on a display, and eventually remember preferences with the user’s permission. The first milestone remains a working conversation loop.
✦
Reason
Send structured requests over HTTPS to a frontier model API. Claude is the initial target reasoning layer.
↗
Respond
Return spoken replies and expressive on-screen feedback. Supported hardware actions can follow once the conversation loop is reliable.
02 / Prototype
A development-board prototype, with a clear next step.
The hardware direction centers on ESP32-S3, audio input and output, and a small display. The next step is to connect this foundation to a Claude-powered backend and evaluate the complete interaction.
DEVICE PLATFORMESP32-S3
Audio input
Display feedback
Wi-Fi connectivity
Prototype direction
Hardware & firmware
- ESP32-S3 development board
- USB microphone input
- 2.2-inch SPI TFT display
- I²S audio amplifier + speaker
- Wi-Fi / HTTPS connectivity
- FreeRTOS-based device logic
Development status
01
Hardware learning & assembly
Earlier independent-study work focused on ESP32 hardware assembly, wiring, soldering, power integration and firmware flashing.
↻
Cloud reasoning integration
Planned: a backend that sends transcribed requests to Claude and returns conversational replies and structured device commands.
→
Productization
Future: a repeatable demo, measured latency and reliability, then enclosure design and onboarding.
03 / Architecture
From a voice input to a physical response.
Proposed architecture: the device handles hardware interaction; a backend coordinates speech services, conversational context and model requests.
1User voiceNatural input
→
2MicrophoneAudio capture
→
3ESP32-S3Device control
→
4Speech / promptTranscription + context
→
5Claude APITarget reasoning layer
→
6OutputsVoice · UI · actions
Integration note: Claude integration is planned and is not yet deployed. Speech recognition and synthesis are separate services. API credentials belong on the backend, rather than in device firmware.
04 / Use cases
A companion for everyday curiosity.
Everyday conversation
A small physical presence for questions, casual conversation and shared curiosity, with voice and display feedback.
Learning together
Explore ideas and explain concepts through dialogue. The hardware itself remains a hands-on embedded-systems learning project.
Expressive interaction
Use an animated face and clear device states to give conversations a physical character. Longer-term personalization is a future goal.
Why Claude
A reasoning layer for interactions that go beyond chat.
We plan to use Claude to interpret transcribed requests, maintain conversational context and generate structured responses. Device actions will be limited to supported commands. Our evaluation will focus on response quality, latency, reliability and cost per interaction.
Context-aware conversationStructured responsesTool / device-action workflowsRapid prototype iteration
05 / Development
From independent study to a focused hardware project.
EdgeTalk grows out of an earlier “build an intelligent conversation device” independent-study project. The emphasis was hands-on embedded learning: assembling hardware, integrating power and loading firmware.
Learn by building
The independent-study work explored ESP32-based conversation hardware and the Xiaozhi open-source project. It covered wiring, soldering, power integration and firmware flashing.
This is the project's learning foundation, rather than an original claim over the open-source firmware.
Connect the full conversation loop
Build a backend that coordinates transcription, Claude requests and speech output. Show a complete voice-to-response interaction on development hardware.
Claude integration is planned. No production deployment or performance figures are claimed.
Measure before productizing
Record response latency, successful interactions, connection failures and API cost. Use the findings to guide device-state feedback, error recovery and enclosure design.
These are proposed evaluation criteria; results will be published after testing.
Engineering priorities
Reliable power and connections • clear listening and response states • recoverable network errors • backend-managed credentials • bounded device actions
06 / About
Independent, early-stage, built in Taiwan.
EdgeTalk Labs is an independent, founder-led hardware project in Taiwan, exploring voice interfaces for embedded devices. The work connects an earlier independent-study effort with a new cloud reasoning integration plan.
The project is at the personal-prototype stage. Its long-term goal is an expressive physical AI companion; the immediate goal is a repeatable voice-to-response demonstration on development hardware. EdgeTalk Labs is a working project name and is not currently an incorporated company.
Contact
Email EdgeTalk
Talk companion hardware, conversation or a future pilot.
For technical questions, feedback or collaboration, contact the project directly.
hello@fatevault.org