Preview

Realtime / WebSocket API (preview)

Low-latency voice streaming over WebSocket. Bidirectional audio with sub-300ms round-trip for conversational AI experiences.

Overview

The Realtime API provides a persistent WebSocket connection for streaming raw PCM audio to and from Meridian models. Unlike the REST transcription endpoint, the realtime path eliminates HTTP overhead per chunk — audio frames flow continuously, and responses stream back as they are generated. This is the same transport used by the Meridian desktop client for its voice mode.

Endpoint

wss://api.getnimbus.net/v1/realtime

TLS 1.3 only. Authentication via Authorization: Bearer <api_key> passed as a query parameter during the WebSocket upgrade handshake.

Protocol

Messages are JSON-framed binary. Each frame carries a type discriminator, a sequence number for ordering, and an optional payload. Audio payloads are 16-bit mono PCM at 24 kHz.

{
  "type": "audio_in",
  "seq": 42,
  "payload": "<base64>"
}

Client → Server

audio_in — PCM chunk
config — model, language, vad settings
ping — keepalive

Server → Client

transcript — partial or final text
audio_out — TTS PCM
error — fault frame
pong — keepalive ack

Roadmap

✓

WebSocket transport

Persistent connection, JSON framing, base64 PCM payloads.

✓

Voice activity detection

Server-side VAD with configurable thresholds.

—

TTS audio out

Streaming text-to-speech responses back over the same socket.

—

Function calling

Tool-use frames for agentic voice workflows.

—

SDK packages

Python & Node.js client libraries with automatic reconnect.

Rate Limits

Tier	Concurrent connections	Audio minutes / hour
Free	1	10
Pro	5	300
Enterprise	Custom	Custom

This API is in active development. The protocol may change before general availability. Join the Meridian Discord for early access and protocol updates.