Realtime providers

A realtime provider is a single speech-to-speech model that takes audio in and streams audio out — no separate STT or TTS. In Syrinx you wrap one in a RealtimeBridge, which runs it as a plugin and gives you the same tools, resume, and observability surface as a cascade.

import { RealtimeBridge, fromOpenAIRealtime } from '@kuralle-syrinx/realtime';

session.registerPlugin('realtime', new RealtimeBridge(fromOpenAIRealtime({ apiKey, socketFactory })));

Run the session with endpointingOwner: "timer" — a realtime model owns its own turn detection, so you don’t register STT/VAD/TTS plugins.

OpenAI Realtime

import { fromOpenAIRealtime } from '@kuralle-syrinx/realtime';

const adapter = fromOpenAIRealtime({
  apiKey: process.env.OPENAI_API_KEY!,
  socketFactory: createNodeWsSocket,
  model: 'gpt-realtime',
});

Option	Notes
apiKey	Required.
model	The OpenAI Realtime model id.
socketFactory	createNodeWsSocket on Node, createWorkersSocket on the edge.
tools	Function tools the model can call — see delegating to a reasoner.

On resume, OpenAI replays the transcript into the model (conversation.item.create) so a hibernated call picks back up with full context.

Gemini Live

import { fromGeminiLive } from '@kuralle-syrinx/realtime';

const adapter = fromGeminiLive({
  apiKey: process.env.GEMINI_API_KEY!,
  socketFactory: createNodeWsSocket,
  systemInstruction: 'You are a helpful voice assistant.',
});

Option	Notes
apiKey	Required.
model	The Gemini Live model id.
systemInstruction	System prompt.
tools	Function tools.

Gemini Live resumes natively from its own session handle — Syrinx passes the handle through rather than replaying the transcript, so there’s no double-applied history.

Grok

Grok’s realtime API is OpenAI-compatible, so it uses the same adapter shape:

import { fromGrokRealtime } from '@kuralle-syrinx/grok';

const adapter = fromGrokRealtime({ apiKey: process.env.XAI_API_KEY!, socketFactory });

Half-cascade: realtime reasoning, your own TTS voice

If you want a realtime front’s reasoning but a specific TTS voice or language, run the front text-only and let a Syrinx TTS plugin speak its transcript. Set the text-only modality on the adapter and register a TTS plugin instead of relying on the front’s audio out. This pairs a strong reasoning front with the TTS provider of your choice.