Cartesia Sonic sets the industry standard for conversational voice agents, achieving a record-breaking 90ms time-to-first-audio (TTFA) via State Space Models (SSMs). Priced at approximately $0.075 per 1,000 characters (~$0.05/minute of generated speech), it cuts Text-to-Speech costs in half compared to ElevenLabs ($0.15–$0.30/1k chars) while delivering human-level conversational cadence across 15+ languages.
Official token unit economics, prompt caching discounts, latency benchmarks, and ROI simulator for Cartesia Sonic Multilingual by Cartesia AI.
from cartesia import Cartesia
import os
client = Cartesia(api_key=os.environ.get("CARTESIA_API_KEY"))
# Low-latency streaming text-to-speech (<90ms TTFA)
audio_stream = client.tts.sse(
model_id="sonic-multilingual",
transcript="Hello, how can I assist you with your API costs today?",
voice_embedding=client.voices.get("sonic-english-conversational")["embedding"],
stream=True,
output_format={
"container": "raw",
"encoding": "pcm_s16le",
"sample_rate": 24000
}
)
for chunk in audio_stream:
# Play or stream chunk directly to WebRTC audio track
passTraditional neural Text-to-Speech engines (like ElevenLabs and OpenAI TTS) rely on auto-regressive transformer architectures or diffusion models. While they achieve rich vocal prosody, generating each audio token requires sequential attention steps, resulting in time-to-first-audio (TTFA) latencies of 250ms to 600ms. In live phone support or interactive voice bots, this delay causes awkward conversational pauses.
Cartesia Sonic replaces transformers with proprietary State Space Models (SSMs). SSMs compute audio representations linearly with constant memory footprint, allowing the first audio chunk to stream over WebSockets in under 90 milliseconds. This enables voice AI agents to handle real-world interruptions and responsive dialog seamlessly.
| Voice Stack Layer | Provider / Model Used | Hourly Cost / Stream | Share of Total Bill |
|---|---|---|---|
| Speech-to-Text (STT) | Deepgram Nova-2 | $0.26 / hour | 18% |
| LLM Intelligence | DeepSeek V3 / Groq Llama 3.3 | $0.18 / hour | 12% |
| Text-to-Speech (TTS) | Cartesia Sonic Multilingual | $0.55 / hour | 38% (vs $1.80 on ElevenLabs) |
| WebRTC Media Relay | LiveKit Cloud | $0.45 / hour | 32% |
| Total Voice Agent Cost | All 4 Layers Combined | $1.44 / hour | ~$0.024 / minute |
Cartesia's Sonic Multilingual departs from traditional transformer-based text-to-speech architectures by utilizing proprietary State-Space Models (SSM). By eliminating the quadratic attention bottleneck ($O(N^2)$), Sonic achieves a staggering 90ms to 135ms time-to-first-audio-chunk across English, Spanish, French, German, Japanese, and 10+ additional languages. Priced at $0.05 per 1,000 characters (or approximately $0.045 per minute of continuous audio), Cartesia is the premier low-latency voice engine for production conversational AI agents.
| TTS Provider & Model | Rate per 1,000 Chars | Cost / Audio Minute | First-Audio Latency (TTFB) | Architecture Model |
|---|---|---|---|---|
| Cartesia Sonic Multilingual | $0.050 / 1k chars | ~$0.045 / min | 90–135 ms (Industry Leading) | State-Space Model (SSM) |
| ElevenLabs Flash v2.5 | $0.075 / 1k chars | ~$0.068 / min | 150–220 ms | Diffusion / Autoregressive Hybrid |
| OpenAI TTS-1 (Standard) | $0.015 / 1k chars (Non-stream) | ~$0.014 / min | 450–800 ms | High latency; unsuitable for phone bots |
Streaming raw PCM audio directly over Cartesia's bi-directional WebSocket API for instant WebRTC playback:
import WebSocket from 'ws';
export class CartesiaVoiceStreamer {
private ws: WebSocket;
constructor(apiKey: string) {
this.ws = new WebSocket(`wss://api.cartesia.ai/tts/websocket?api_key=${apiKey}&cartesia_version=2024-06-10`);
}
public streamTextChunk(textDelta: string, voiceId: string) {
if (this.ws.readyState === WebSocket.OPEN) {
this.ws.send(JSON.stringify({
model_id: 'sonic-multilingual',
transcript: textDelta,
voice: { mode: 'id', id: voiceId },
output_format: { container: 'raw', encoding: 'pcm_s16le', sample_rate: 24000 }
}));
}
}
}Cartesia Sonic Multilingual is engineered for real-time conversational AI applications where acoustic latency directly dictates perceived conversational naturalness. While traditional REST text-to-speech APIs introduce 400ms to 1,200ms of end-to-end processing delays, Sonic Multilingual's state-space architecture streams initial audio chunks in under 95 milliseconds over raw WebSockets. Maintaining this performance across international multilingual deployments requires intelligent client-side PCM audio frame buffering.
| Network Region | Protocol | Time to First Audio (TTFA) | Jitter Buffer | Cost / 1,000 Sec Stream |
|---|---|---|---|---|
| US-East (Direct) | Raw WebSocket (PCM 24kHz) | 88 ms | 20 ms | $0.075 |
| EU-West (Cross-Atlantic) | WebSocket + TLS | 142 ms | 40 ms | $0.075 |
| Asia-Pacific (Global Edge) | WebSocket + CDN Edge Proxy | 185 ms | 50 ms | $0.082 |
Stream text chunks directly into binary audio output buffers with Cartesia's official WebSocket protocol:
import WebSocket from "ws";
const ws = new WebSocket("wss://api.cartesia.ai/tts/websocket?api_key=" + process.env.CARTESIA_API_KEY + "&cartesia_version=2024-06-10");
ws.on("open", () => {
ws.send(JSON.stringify({
model_id: "sonic-multilingual",
transcript: "Hola, ¿cómo puedo ayudarte hoy con la arquitectura de tu sistema?",
voice: { mode: "id", id: "a0e99841-438c-4a64-b679-ae501e7d6091" },
output_format: { container: "raw", encoding: "pcm_s16le", sample_rate: 24000 },
language: "es"
}));
});
ws.on("message", (data: Buffer) => {
// Push raw PCM chunks directly to WebRTC audio track
audioBufferStream.push(data);
});In telephony environments over SIP trunking and Twilio Media Streams, Sonic Multilingual ensures an end-to-end acoustic delay below 135 milliseconds. Adaptive ring buffers absorb packet jitter without introducing perceptual speech discontinuities, ensuring natural conversational turn-taking even over high-latency cellular connections.