← all demos
hospitality · surface: front-desk-interpreter
Front desk interpreter
Two languages, one front desk. Real-time, both directions. Hotels and front desks that serve international guests without bilingual staff on every shift.
NO LIVEKIT CREDENTIALSPaste your LiveKit URL, key and secret to connect the scope to a live room.
CANVAS · front-desk-interpreterwaiting
agent-mounted UIappears here on connect
Two languages, one front desk. Real-time, both directions. Hotels and front desks that serve international guests without bilingual staff on every shift.
agent may draw
CardCaptions
REQUIRED CREDENTIALS
livekit_urllivekit_api_keylivekit_api_secretLiveKit: ✗ add your keys
Run the worker locally:then connect: the agent joins room front-desk-interpreter.
uv run python agent.py devbuild writeup
The leanest agent in the cookbook is a hotel interpreter
by Mahimai
The problem
A hotel front desk gets a guest who speaks no English. Today that means a phone tree, a paid interpreter line, or a clerk and guest poking at translation apps. The job is small and well defined: hear one side, say it back to the other, do nothing else.
Why this stack
One model does the whole thing. Gemini Live is speech to speech, so there is no separate STT, no TTS, no VAD model, and no turn detector to wire up or pay for. The agent is the leanest in the cookbook: an instructions string, a session, and two event handlers for captions. English anchors the desk side; the guest language is whatever the model last heard that was not English, so direction lives in the prompt, not in code.
The interesting part
The session is the whole runtime. No plugins beyond the realtime model:
session = AgentSession(
llm=google.realtime.RealtimeModel(
model=GEMINI_LIVE_MODEL,
voice="Puck",
input_audio_transcription=types.AudioTranscriptionConfig(),
output_audio_transcription=types.AudioTranscriptionConfig(),
),
)
Captions pair the guest line with its translation: user_input_transcribed
accumulates multi-segment finals (one long utterance emits several), and
conversation_item_added flushes the pending original onto the assistant
row.
What surprised me
The model id is load-bearing and the docs lie about the default. Passing
gemini-2.5-flash (a text-API id) makes the Live websocket close with a
1008 policy violation: that model has no bidiGenerateContent. The id
that works for a Gemini API key is
gemini-2.5-flash-native-audio-preview-12-2025. An afternoon went to a
one-line constant.
Run it
Talk to it at https://playground.mahimai.ca/demos/front-desk-interpreter. Open in English, switch to any other language mid-call, and hand the phone back and forth. Or fork the cookbook and run the worker locally.