Phase 3: Text-to-Speech (TTS) · 55 min · Python · ElevenLabs SDK · Cartesia SDK

ElevenLabs and Cartesia Integration

ElevenLabs for quality, Cartesia for speed — knowing when to use each is what separates senior voice AI engineers from juniors.

Hiring signal: TTS provider integration with TTFB measurement is a standard coding exercise in voice AI interviews at Vapi and Retell.

What you will learn

Integrate ElevenLabs for high-quality TTS with voice selection and streaming
Integrate Cartesia Sonic-3 for low-latency streaming TTS via WebSocket
Measure and compare TTFB between ElevenLabs and Cartesia
Choose between ElevenLabs and Cartesia based on use case requirements

The Problem

ElevenLabs and Cartesia are the two most-used TTS providers in production voice agents. They have different APIs, different streaming models, different pricing, and different strengths. You need to integrate both — ElevenLabs for premium experiences, Cartesia for latency-critical and cost-sensitive workloads.

The Concept

Provider Comparison

Feature	ElevenLabs	Cartesia Sonic
Architecture	Autoregressive	Flow-Matching
TTFB	~400ms	~120ms
MOS	4.5	4.2
Price/min	$0.30 (Tier 1)	$0.015
Voice cloning	Yes (instant + professional)	Yes (limited)
Streaming	WebSocket + HTTP chunked	WebSocket
Languages	29	15
Sample rates	44.1kHz, 32kHz	44.1kHz, 24kHz, 16kHz
Emotion control	Yes (via voice settings)	Yes (via context)
Speed control	Yes	Yes
API	REST + WebSocket	WebSocket

ElevenLabs API

from elevenlabs import ElevenLabs

client = ElevenLabs(api_key=os.getenv("ELEVENLABS_API_KEY"))

# Streaming synthesis
audio_stream = client.text_to_speech.stream(
    text="Hello, how can I help you today?",
    voice_id="21m00Tcm4TlvDq8ikWAM",
    model_id="eleven_turbo_v2_5",
    voice_settings={"stability": 0.5, "similarity_boost": 0.75},
)

for chunk in audio_stream:
    # chunk is raw audio bytes (MP3 or PCM)
    yield chunk

Cartesia API

import websockets
import json

async def cartesia_stream(text, voice_id="c45a6b5a-9c3b-4c3b-8c3b-4c3b8c3b4c3b"):
    url = "wss://api.cartesia.ai/tts/websocket"
    headers = {"Authorization": f"Bearer {os.getenv('CARTESIA_API_KEY')}"}

    async with websockets.connect(url, extra_headers=headers) as ws:
        msg = {
            "transcript": text,
            "voice": {"id": voice_id},
            "output_format": "pcm_16000",
            "model": "sonic-2",
        }
        await ws.send(json.dumps(msg))

        while True:
            response = await ws.recv()
            data = json.loads(response)
            if data.get("type") == "chunk":
                yield data["data"]  # Audio bytes
            elif data.get("type") == "done":
                break

You're building a voice agent that needs both premium quality for VIP callers and low-latency for standard callers. How do you handle TTS provider selection?

Always use Cartesia for speed

Dual-Provider Routing

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers Build It, Use It, Ship It, Evaluation, Key Terms, Common Pitfalls, Interview Framing — plus a hands-on lab, quiz, and project artifact.

Create a free account to unlock Phase 0 and Phase 1 of every course — no credit card.