Front desk interpreter

A hotel front desk gets a guest who speaks no English. The job is small and well defined: hear one side, say it back to the other, do nothing else. This agent does exactly that, both directions, in real time.

One model does the whole thing. Gemini Live is speech to speech, so there is no separate STT, no TTS, no VAD, and no turn detector. The session is the whole runtime:

session = AgentSession(
    llm=google.realtime.RealtimeModel(
        model=GEMINI_LIVE_MODEL,
        voice="Puck",
        input_audio_transcription=types.AudioTranscriptionConfig(),
        output_audio_transcription=types.AudioTranscriptionConfig(),
    ),
)

The desk-side language is not hard-coded. The host picks it in the playground and it rides the agent-dispatch metadata, arriving as ctx.job.metadata. The agent defaults to English and caps the length, since a browser-minted value is untrusted.

The catch is that a single session links to one participant, the first to join. On a two-party call it would hear only one side. So the agent tracks the active speaker and re-links the session to whoever is talking, which makes it interpret both ways:

current = getattr(room_io, "linked_participant", None)
if current is None or current.identity != sp.identity:
    room_io.set_participant(sp.identity)

Captions pair each line with its translation. user_input_transcribed accumulates the guest's speech (one long utterance emits several final segments), and conversation_item_added flushes that pending original onto the assistant row when the reply lands:

row: dict = {"text": text}
if pending["original"]:
    row["original"] = pending["original"]
    pending["original"] = None
captions.append(row)

One more multiparty detail: the invited guest joins after the initial UI broadcast, so it asks for the scene and captions once its data handler is listening, and the agent replays them to just that participant.

The model id is load-bearing and a text-API id will not work. Passing gemini-2.5-flash closes the Live socket with a 1008 policy violation: that model has no bidiGenerateContent. Use gemini-2.5-flash-native-audio-preview-12-2025.

Build it from an empty folder in the full walkthrough, or talk to the finished agent at https://playground.mahimai.ca/demos/front-desk-interpreter.