EMO Server

Custom AI server for the EMO robot. Replaces living.ai cloud with local AI processing.

Architecture

EMO robot (192.168.1.154)
       │
       ▼
nginx (SSL termination, port 443)
       │
       ├── /tts/dl/*  ──────────►  emo-ai (Python, port 9090)
       │                              - serves cached audio files
       │                              - prefers _emovoice.mp3 version
       │
       └── everything else  ────►  emo-proxy (Go, port 8080)
                                      │
                                      ├── /emo/ai/detectintent
                                      │     sends audio to emo-ai /process
                                      │     emo-ai: Vosk STT → triggers → cache → GPT → TTS
                                      │     returns response to EMO
                                      │     background: calls living.ai TTS, saves _emovoice.mp3
                                      │
                                      ├── /emo/ai/imgrecog
                                      │     saves photo to /home/homer/emo-photos/
                                      │     proxies to living.ai for ChatGPT vision response
                                      │
                                      └── all other EMO API calls
                                            proxied to living.ai as-is

DNS (dnsmasq): all living.ai domains → 192.168.1.64

How it works

1. EMO sends voice audio to api.living.ai (DNS redirected to our server)
2. nginx terminates SSL, routes to emo-proxy
3. emo-proxy forwards audio to emo-ai /process endpoint
4. emo-ai pipeline:
   • Converts 16-bit big-endian PCM → WAV (byte-swap)
   • Vosk STT (Russian, large model ~2.8GB) transcribes speech
   • Checks voice triggers (dance, zombie, show_something, etc.)
   • Checks response cache (SHA256 of text)
   • If no cache hit → GPT-4o-mini generates response
   • OpenAI TTS generates audio (alloy voice)
   • Saves audio as {audio_id}.mp3, returns response JSON
5. emo-proxy returns response to EMO immediately
6. Background: proxy calls living.ai TTS with the same text, saves as {audio_id}_emovoice.mp3
7. Next time EMO requests the same audio → emo-ai serves _emovoice.mp3 (EMO's own voice)

Project structure

proxy/                    Go reverse proxy
  emoProxy.go             main proxy code (~1100 lines)
  emoProxy.conf.example   config template (triggers, API servers)
  Dockerfile              builds the proxy container
  go.mod, go.sum          Go dependencies

ai-server/                Python AI server
  server.py               main AI server (~700 lines)
  Dockerfile              builds the AI container (Vosk, OpenAI, ffmpeg)

nginx/                    nginx config
  default.conf            SSL + routing rules

deploy/                   deployment configs
  docker-compose.yml      3 containers: nginx, emo-proxy, emo-ai
  dnsmasq.conf            DNS redirects for living.ai domains
  deploy.sh               manual deploy script
  autodeploy.sh           auto-deploy (runs via cron)
  .env                    secrets (OPENAI_API_KEY) — not in git

Setup from scratch

1. DNS: configure router or dnsmasq to redirect living.ai domains to your server IP
   • domains: api.living.ai, eu-api.living.ai, eu1-api.living.ai, eu-tts.living.ai, res.living.ai, res-eu.living.ai, us-api.living.ai, as-api.living.ai
2. SSL: generate certs for *.living.ai → place in nginx/ssl/
3. Config: copy proxy/emoProxy.conf.example → proxy/emoProxy.conf
4. Secrets: create deploy/.env with OPENAI_API_KEY=sk-...
5. Run: cd deploy && docker compose up -d

Auto-deploy

Every 2 minutes, a cron job on the server runs deploy/autodeploy.sh:

*/2 * * * * /home/homer/emo-server/deploy/autodeploy.sh >> /home/homer/autodeploy.log 2>&1

The script:

1. git fetch origin main — checks for new commits
2. Compares local HEAD vs remote HEAD
3. If different → git pull
4. If proxy/emoProxy.go changed → copies to /home/homer/Proxy/, rebuilds Docker image, restarts emo-proxy
5. If ai-server/server.py changed → copies to /home/homer/emo-ai-server/, restarts emo-ai
6. Logs everything to /home/homer/autodeploy.log

Workflow: edit code → git push → within 2 minutes the server pulls and deploys automatically.

Voice options

Engine	Speed	Quality	Internet	Notes
OpenAI TTS	~1s	excellent	yes	current default
living.ai TTS	~3s	original EMO	yes	background caching
Piper TTS	~400ms	good	no	4 Russian voices installed
espeak-ng	~50ms	robotic	no	classic robot voice

Piper voices installed at /home/homer/piper-voices/:

• ru_RU-denis-medium — male
• ru_RU-dmitri-medium — male
• ru_RU-irina-medium — female
• ru_RU-ruslan-medium — male

API endpoints (emo-ai, port 9090)

Endpoint	Method	Description
/process	POST	Main pipeline: audio → STT → GPT → TTS
/action	POST	Queue action: {"action": "dance"}
/say	POST	Queue TTS phrase: {"text": "Привет!"}
/health	GET	Health check
/actions	GET	List supported actions
/triggers	GET	List voice triggers
/tts/dl/{id}	GET	Serve cached audio file

Useful commands

# Queue action
curl -X POST http://192.168.1.64:9090/action -H 'Content-Type: application/json' -d '{"action":"zombie"}'

# Queue speech
curl -X POST http://192.168.1.64:9090/say -H 'Content-Type: application/json' -d '{"text":"Привет!"}'

# Restart services
docker restart emo-proxy emo-ai

# Check autodeploy log
tail -f /home/homer/autodeploy.log

# Manual deploy
/home/homer/emo-server/deploy/deploy.sh