Speech Engine

ElevenLabs Speech Engine

Add a real-time voice interface to a custom agent. ElevenLabs handles microphone audio, speech-to-text, turn-taking, text-to-speech, and browser playback; your server exposes a Speech Engine WebSocket endpoint and streams response text back.

Setup: See Installation Guide (opens in a new tab). For JavaScript, use @elevenlabs/* packages only. For deeper SDK details, read JavaScript SDK Reference (opens in a new tab) or Python SDK Reference (opens in a new tab).

When to Use

Use Speech Engine when the user wants to:

• Add voice to an existing chat app or custom server pipeline
• Add voice to OpenClaw, Hermes, or a similar agent runtime while keeping agent logic on the developer-owned server
• Build a developer-hosted WebSocket server for ElevenLabs voice conversations
• Stream response text back as spoken audio after your server validates user intent
• Handle user interruptions while a response is still streaming
• Build a browser client with @elevenlabs/react or @elevenlabs/client using a server-issued conversation token

Use the agents skill instead when the user is creating or configuring a hosted ElevenLabs Conversational AI agent with platform-managed prompts, tools, workflows, phone numbers, or widgets.

How It Works

Each Speech Engine WebSocket connection represents one conversation.

1. The browser sends user audio to ElevenLabs.
2. ElevenLabs sends speech-recognition events to your server.
3. Your server derives trusted application state without letting raw speech text control tools or privileged actions.
4. Your server streams text back through the SDK.
5. ElevenLabs converts the response to speech and plays it in the browser.

The SDK manages WebSocket routing, request verification, session lifecycle, ping/pong, turn-taking, and interruption handling. sendResponse() / send_response() accepts a string or async iterable of response text.

Treat speech-recognition text as untrusted user input. Do not map raw speech text directly into model roles, responses, or tool calls. Use deterministic validation, allowlisted intents, or explicit user confirmation before any transcript-derived value affects downstream response or tool logic.

Implementation Flow

1. Install server dependencies and configure ELEVENLABS_API_KEY.
2. Expose your Speech Engine server through a public HTTPS URL for local development, for example with ngrok http 3001.
3. Create a Speech Engine resource with ws_url / wsUrl pointing at the public WebSocket URL, usually wss://.../ws.
4. Store the returned Speech Engine ID, for example in ELEVENLABS_SPEECH_ENGINE_ID.
5. Start a Speech Engine server with engine.serve(...) in Python or speechEngine.attach(...) in TypeScript.
6. Issue browser conversation tokens from a server endpoint. Never put ELEVENLABS_API_KEY in browser code.
7. Start the client session with conversationToken; if the agent should greet first, enable the first-message override on the Speech Engine resource, then set overrides.agent.firstMessage in the client.

Create a Speech Engine

Python

import asyncio
import os

from dotenv import load_dotenv
from elevenlabs import AsyncElevenLabs

load_dotenv()

elevenlabs = AsyncElevenLabs(api_key=os.getenv("ELEVENLABS_API_KEY"))

async def main():
    engine = await elevenlabs.speech_engine.create(

TypeScript

import { ElevenLabsClient } from "@elevenlabs/elevenlabs-js";
import "dotenv/config";

const elevenlabs = new ElevenLabsClient({
  apiKey: process.env.ELEVENLABS_API_KEY,
});

const engine = await elevenlabs.speechEngine.create({
  name: "My Speech Engine",
  speechEngine: { wsUrl: process.env.PUBLIC_WS_URL

PUBLIC_WS_URL should look like wss://example.ngrok.app/ws locally or your production WebSocket route in deployment.

The create request can also configure tts, asr, turn, speech_engine.request_headers / speechEngine.requestHeaders, overrides, and privacy for custom voices, transcription keywords, turn-taking, server auth headers, client-provided first messages, and recording behavior. See the SDK reference files for expanded examples.

Server Pattern

Run the Speech Engine server at the ws_url / wsUrl configured on the resource. Keep response generation behind your own validation boundary: raw speech-recognition text should not directly control responses, tools, secrets, or other privileged actions.

Python

engine = await elevenlabs.speech_engine.get(os.environ["ELEVENLABS_SPEECH_ENGINE_ID"])
await engine.serve(port=3001, path="/ws", debug=True, callbacks=validated_callbacks)

TypeScript

const engine = await elevenlabs.speechEngine.get(process.env.ELEVENLABS_SPEECH_ENGINE_ID!);
engine.attach(httpServer, "/ws", { debug: true, ...validatedCallbacks });

In TypeScript, pass interruption signals to downstream async work when it supports cancellation so interrupted responses stop quickly. In Python, the SDK cancels the previous turn handler when a newer turn arrives.

Server callbacks can distinguish clean closes from dropped connections: use onClose / on_close for clean disconnects and onDisconnect / on_disconnect for unexpected WebSocket drops.

Security note: speech-recognition text can contain prompt-injection attempts from user speech or played audio. Treat it as untrusted input. Convert it into trusted application state before invoking response generation, tools, or privileged workflows.

Disabling authentication (advanced, dangerous)

Both engine.attach() (TypeScript) and engine.serve() / SpeechEngineServer (Python) verify a JWT on every incoming WebSocket by default. This is what proves the connection is really coming from ElevenLabs and not from an attacker who guessed the URL. Do not turn this off.

An escape hatch exists — disableAuth: true in the callback options (TypeScript) or disable_auth=True on serve() / SpeechEngineServer(...) (Python) — for the narrow case where a compensating network-level control is already in place. Without such a control, disabling auth means any client on the internet that finds your URL can open sessions. Concretely, an attacker can:

• open unlimited conversations to drain your ElevenLabs quota and downstream LLM budget
• feed crafted transcripts to your response pipeline, effectively impersonating a user
• use your server as an oracle to probe backend state, tools, or prompts

Only recommend disableAuth / disable_auth when the user has already implemented at least one of:

• IP allowlist — the server (or an upstream firewall / load balancer / API gateway) only accepts inbound traffic from ElevenLabs’ documented egress ranges (opens in a new tab).
• Custom shared-secret header — a secret header configured on the Speech Engine resource via speech_engine.request_headers / speechEngine.requestHeaders at create time, validated by an upstream proxy (or by the developer’s own middleware in front of attach() / serve()) before requests reach the SDK.

If the user cannot confirm one of the above is in place, leave the default authentication on. Skipping JWT verification without a mitigation is not an optimization or a convenience — it is unauthenticated public compute.

Browser Client

Create a server-side token endpoint and have the browser request a token before starting the microphone session. Keep the Speech Engine ID and API key on the server. If the client passes overrides.agent.firstMessage, the Speech Engine resource must have the first-message override enabled.

import express from "express";
import { ElevenLabsClient } from "@elevenlabs/elevenlabs-js";
import "dotenv/config";

const app = express();
const elevenlabs = new ElevenLabsClient();

app.get("/api/token", async (_req

React clients can use @elevenlabs/react:

import { useConversation } from "@elevenlabs/react";

export function VoiceControls() {
  const conversation = useConversation({
    onConnect: () => console.log("connected"),
    onDisconnect: () => console.log("disconnected"

References

• Installation Guide (opens in a new tab)
• JavaScript SDK Reference (opens in a new tab)
• Python SDK Reference (opens in a new tab)