avatar

# Phonagnosia Voice-first multilingual assistant built for live expo demos. Project streams LLM text into Streamlit chat, transcribes microphone input with Siemens STT, calls MCP tools when available, and synthesizes spoken responses with ElevenLabs in same detected language. Concrete use case: expo visitor asks about Siemens GBS CEE FPS capabilities in English, Czech, or German, and assistant answers in same language as spoken holographic colleague. ## Why this repo exists This started as fast-turnaround expo prototype for holographic AI assistant demo. Repo now cleaned up to show practical full-stack AI integration work: real-time UX, voice I/O, prompt design, tool calling, and external service orchestration. ## Demo flow 1. Visitor types or records message in Streamlit UI. 2. Audio input goes through Siemens Whisper transcription. 3. Text service detects language (`en`, `cs`, `de`) and streams response from Siemens-hosted LLM. 4. LLM can call local MCP tools over `streamable-http`. 5. Response is converted to speech with ElevenLabs and played back in UI. ## Stack - Python 3.13+ - `uv` for environment and dependency management - Streamlit for chat UI - Siemens LLM/STT endpoints via OpenAI-compatible SDK - FastMCP for local tool server - ElevenLabs for text-to-speech ## Quick start ```bash uv sync cp config/.env.example config/.env make start ``` Open Streamlit URL shown in terminal. `make start` launches local MCP server and app together. ## Required environment variables Set these in `config/.env`: ```bash SIEMENS_API_KEY=... ELEVENLABS_API_KEY=... ``` Optional override: ```bash MCP_BASE_URL=http://localhost:8000/mcp ``` ## Useful commands ```bash make run-mcp # start local MCP server only make run-app # start Streamlit app only make start # start both for local development make test # run unit tests make test-stt # post sample wav file to Siemens STT endpoint make docker-build make docker-run ``` ## Automation - GitHub Actions CI runs unit tests and Python compile checks on pull requests and pushes to `main`. - GitHub Actions CD runs on `main` and publishes verified source bundle as build artifact. - GitHub Actions image workflow builds and publishes `ghcr.io/<owner>/<repo>` container image on merges to `main`. ## Container run ```bash cp config/.env.example config/.env make docker-build make docker-run ``` Container starts local MCP server and Streamlit app in one process group. App becomes available on `http://localhost:8501`. ## Project structure - `src/app.py`: Streamlit chat UI - `src/server.py`: local MCP server - `src/services/text.py`: language detection, LLM streaming, MCP tool calls - `src/services/speech_to_text.py`: Siemens Whisper wrapper - `src/services/text_to_speech.py`: ElevenLabs streaming TTS - `src/resources/templates/`: system and persona prompts - `config/config.yaml`: runtime config ## Current limitations - Voice input transcription currently assumes English because Siemens STT endpoint in this setup does not auto-detect language. - TTS is generated after full text response is produced; audio playback is not sentence-streamed yet. - Default MCP setup is local demo server with small tool surface. ## Portfolio note Repo intentionally keeps implementation small and readable. Goal is not framework complexity; goal is showing end-to-end integration between multimodal UI, LLM orchestration, and tool-augmented inference.