Moving an OpenAI Realtime agent to Dialog-RSN-1

Companion code for Migrating from OpenAI Realtime.

One voice agent, twice. before.py is a front-desk agent for the fictional Grand Meridian Hotel, written against the OpenAI Realtime API with the official openai SDK. after.py is the same file moved to Dialog-RSN-1. The agent, its tools and its bookings live in hotel.py and don't change.

diff before.py after.py

That diff is the whole migration. Every changed line in after.py has a MIGRATION: comment.

What changes

	before.py (OpenAI)	after.py (Dialog-RSN-1)
Base URL	https://api.openai.com/v1	https://api.us.poly.ai/v1
Key	OpenAI API key	Dialog-RSN-1 workspace API key
Replies	output_modalities: ["audio"], a voice, an audio output format	output_modalities: ["text"], no audio.output
Transcription	gpt-4o-mini-transcribe	none to configure: the model transcribes as part of its turn
Turn detection	server_vad with tuned threshold, padding and silence	server_vad with create_response only
Reply events	response.output_audio.delta and its transcript	response.output_text.delta
Speaker, echo muting, barge-in clearing	yes	removed: there's no assistant audio

Everything else stays: the SDK, client.realtime.connect(), the session shape, tool definitions, function calling, input_audio_buffer.append and the transcription events.

The SDK still sends the key as Authorization: Bearer, and Dialog-RSN-1 accepts that. The model name in connect() is only a label here. The service runs Dialog-RSN-1 whatever you pass.

What happens if you only change the URL and key

Point before.py at Dialog-RSN-1 unchanged and it connects, then prints this and nothing else:

[error] 'audio.output' is not supported by this service. This service has no audio output, so bring your own TTS.

A rejected session.update applies none of its fields, so the instructions and tools go with it. The model still answers, but in text events that before.py never listens for. The same goes for the tuned turn-detection fields: threshold, prefix_padding_ms, silence_duration_ms, interrupt_response and semantic_vad are each rejected with unsupported_parameter.

Run it

You need Python 3.12 and uv.

cp .env.example .env    # add DIALOGUE_API_KEY, and OPENAI_API_KEY to run before.py
uv sync

uv run after.py --say "Can you look up booking GM40912?"   # one text turn
uv run after.py                                            # talk through the microphone

before.py takes the same flags and speaks its replies through your speakers.

To try the audio path without a microphone, make a clip and stream it with --file:

say -o call.aiff "Could you look up my booking? The code is G M four one zero five five."
afconvert -f WAVE -d LEI16@24000 -c 1 call.aiff call.wav
uv run after.py --file call.wav

List microphones with uv run python -m sounddevice and pick one with --device.

Try saying

• "Hi, could you look up my booking? The code is G M four zero nine one two."
• "Can I get a late checkout at three on Thursday?"
• "I need to cancel my reservation. The last name is Reid."

Demo bookings: GM40912 (Okafor), GM41055 (Reid), GM41182 (Reid, a suite), GM39877 (Bergmann, already checked out). Dates are relative to today, and nothing is saved.

Where the voice goes

Dialog-RSN-1 replies in text. To speak the replies, stream response.output_text.delta into a text-to-speech provider. The streaming to text-to-speech guide covers it, including how to report playback so interruptions work across the whole reply.

License

Apache 2.0. See LICENSE.