Guide

Bandwidth voice agents with Cartesia speech on every call

A small server answers Bandwidth's webhook and streams the call's 8 kHz mu-law audio to Cartesia and back without resampling.

Get an API keyRead the docs
server.py
CARTESIA_VERSION = "2026-03-01"
TTS_URL = f"wss://api.cartesia.ai/tts/websocket?cartesia_version={CARTESIA_VERSION}"
# Ink 2 reads the call's native 8 kHz mu-law directly, so caller audio needs no resampling.
STT_URL = (
    "wss://api.cartesia.ai/stt/websocket"
    f"?model=ink-2&cartesia_version={CARTESIA_VERSION}"
    "&encoding=pcm_mulaw&sample_rate=8000&language=en"
)

@app.post("/bxml")
async def bxml() -> Response:
    ws_url = PUBLIC_URL.replace("https://", "wss://") + "/stream"
    body = (
        '<?xml version="1.0" encoding="UTF-8"?>'
        "<Response>"
        f'<StartStream destination="{ws_url}" mode="bidirectional"/>'
        '<Pause duration="600"/>'
        "</Response>"
    )
    return Response(content=body, media_type="application/xml")

Cartesia docs: Bandwidth + Cartesia

What you can do with Bandwidth

What teams build when they put Cartesia and Bandwidth together.

No resampling

Bandwidth carries 8 kHz mu-law and both Cartesia sockets accept pcm_mulaw at 8000 Hz, so audio crosses the bridge as is.

Few dependencies

The example bridge needs only fastapi, uvicorn, httpx, websockets and python-dotenv on Python 3.11 or later.

Your logic in the middle

Your server sits between the two sockets, so the reply text the agent speaks can come from any LLM you choose to call.

Set it up

A documented bridge you run yourself. Bandwidth carries the call and your server connects it to Cartesia's sockets. This is not a Managed Agents number route. Vendor details read in October 2026.

  1. Get the accounts

    A Cartesia API key and voice ID, and a Bandwidth account with Voice API credentials, a number and a Voice Application.

  2. Expose your server

    Use an HTTPS tunnel that supports WebSockets while you develop.

  3. Return BXML

    Answer Bandwidth's webhook with a StartStream verb in bidirectional mode.

  4. Bridge the sockets

    Forward caller audio to the speech to text socket, and speak each final transcript through text to speech.

Documentation

Before you build

Your server, your responsibility

You run the bridge on your own server, so you own its uptime and how it handles the call audio that passes through it.

Want Cartesia to host the agent?

A Managed Agent runs the whole loop for you. Any carrier that sends SIP over TLS or TCP can reach one. See the SIP trunking page.

More integrations

All integrations
SIP trunk

Telnyx

Run a Cartesia voice agent on Telnyx numbers with a SIP trunk, at your Telnyx rates.

  • Managed Agents
See the integration
SIP trunk

Plivo

Run a Cartesia voice agent on Plivo numbers with a Zentrunk SIP trunk, at your Plivo rates.

  • Managed Agents
See the integration
SIP trunk

Vonage

Run a Cartesia voice agent on Vonage numbers with Vonage SIP Trunking, at your Vonage rates.

  • Managed Agents
See the integration

Bandwidth integration FAQs

Can Bandwidth connect to an AI voice agent?

Yes, through a server you run. Bandwidth answers the call, your server returns BXML with a bidirectional StartStream verb, and the server passes the call audio to Cartesia speech to text and speaks replies with Cartesia text to speech. Bandwidth's developer docs have their own Cartesia integration page.

Is this the same as a Managed Agent?

No. This is a bridge you host, with your own LLM calls in the middle. A Managed Agent runs the whole loop on Cartesia. Carrier numbers reach a Managed Agent over a SIP trunk on TLS or TCP, where agents run at $0.03/min, plus LLM usage.

What audio format does Bandwidth stream?

8 kHz mu-law in both directions. Both Cartesia sockets accept pcm_mulaw at 8000 Hz, and Ink 2 reads it directly, so nothing is resampled on the way through the bridge.

What does the Bandwidth bridge cost?

Three separate bills. Bandwidth charges for the call, your LLM provider charges for the replies, and Cartesia charges API credits on your plan. Sonic text to speech costs 1 credit per character and Ink speech to text 3 credits per second of audio. The Free plan includes 20K credits a month, and paid plans start at $5/month.

Can I run the Bandwidth bridge locally while I develop?

Yes. Expose your server through an HTTPS tunnel that supports WebSockets, so Bandwidth can reach your BXML webhook and open the stream. Once it is live, you run the bridge on your own server and own its uptime and how it handles the call audio.

Get started today

Talk to an expert.

Connect with a member of our team and learn how Cartesia can help you build world-class voice experiences.

Contact Sales

Start building.

Access our models via API and bring a voice agent into production in minutes.

Try Cartesia