Intercom

Local-first voice bridge for Arbiter: ESP32 (or any client) speaks PCM in, Intercom runs whisper.cpp STT + Kokoro TTS, and Arbiter stays text + SSE in the middle.

ESP32  --HTTP PTT PCM-->  Intercom  --text/SSE-->  arbiter --api
ESP32  <--chunked PCM---  Intercom  <--text------/

Colocate Intercom on the same host as arbiter --api (default http://127.0.0.1:8080). Device tokens never see the Arbiter bearer.

The default agent is Arthur — a British voice assistant with full Arbiter tool access (config/arthur.agent.json).

Build

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

Requires C++20, CMake 3.20+, SQLite3, Threads. Fetches cpp-httplib and nlohmann/json.

Configure

cp config/intercom.example.json intercom.json
# set arbiter_token, paths to whisper-cli + model, kokoro binary + voice
./build/intercom --config intercom.json

Install speech tools separately (not vendored):

• whisper.cpp → whisper-cli + ggml-base.en.bin
• Kokoro → kokoro-tts + ONNX model and voices bundle

Quick test (no mic)

Fast-path (skips Arbiter):

curl -N -H "Authorization: Bearer dev-device-secret-change-me" \
  -H "X-Device-Id: speaker-1" \
  -H "Content-Type: application/json" \
  -d '{"text":"what time is it"}' \
  --output reply.pcm \
  http://127.0.0.1:8090/v1/utterance/text

Play: ffplay -f s16le -ar 16000 -ac 1 reply.pcm

PCM utterance: see docs/api.md and docs/device.md.

License

Apache-2.0