Cartesia Pricing & Developer Hub

Cartesia Sonic Architecture: State-Space Synthesis & Sub-90ms Audio TTFB

Cartesia powers the next generation of conversational voice AI with its proprietary Sonic State-Space Model (SSM). Traditional autoregressive text-to-speech architectures (like ElevenLabs or StyleTTS) require buffering text clauses and executing heavy transformer attention steps, resulting in 250ms to 450ms of audio generation latency. Cartesia Sonic uses linear-time state-space representations that generate the first streaming audio packet in under 85 milliseconds. At a flat rate of $0.045 per 1,000 characters (~$0.27 per audio hour), Cartesia provides superior voice naturalness at 1/4th the price of legacy TTS providers.

Cartesia Sonic Voice Pricing & Latency Comparison

Low-Latency WebSocket Streaming Client (TypeScript)

import WebSocket from 'ws';

export class CartesiaStreamer {
  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 streamChunk(contextId: string, textChunk: string, onAudio: (pcm: Buffer) => void) {
    this.ws.send(JSON.stringify({
      context_id: contextId,
      model_id: 'sonic-english',
      transcript: textChunk,
      voice: { mode: 'id', id: 'a0e99841-438c-4a64-b679-ae501e7d6091' },
      output_format: { container: 'raw', encoding: 'pcm_s16le', sample_rate: 16000 }
    }));

    this.ws.on('message', (data: string) => {
      const parsed = JSON.parse(data);
      if (parsed.data && parsed.context_id === contextId) {
        onAudio(Buffer.from(parsed.data, 'base64'));
      }
    });
  }

  public cancelStream(contextId: string) {
    // Instantaneous barge-in cancellation: Purges server-side synthesis queue
    this.ws.send(JSON.stringify({ context_id: contextId, cancel: true }));
  }
}

Production SRE Guardrails

• Continuous WebSocket Heartbeats: Cartesia closes idle WebSocket connections after 60 seconds of inactivity. In conversational telephony applications, send periodic WebSocket ping frames to keep the TCP/TLS socket warm between customer speaking turns.
• Sample Rate Alignment: Request 16,000 Hz or 8,000 Hz `pcm_s16le` directly from Cartesia when interfacing with PSTN telephone networks (Twilio/Telnyx). Requesting 44,100 Hz audio wastes 3x server bandwidth and requires client-side downsampling CPU cycles.

Engineering Q&A: Cartesia Voice Infrastructure

Adaptive Jitter Buffering & Telephony Audio Framing (PCM16)

Integrating Cartesia Sonic into carrier-grade telephony networks (FreeSWITCH, Asterisk, or LiveKit SIP) requires precise audio packetization. Cellular networks introduce packet arrival jitter that can cause audio stutter. By configuring an adaptive 40ms jitter buffer and streaming raw 20ms linear PCM16 frames at 8,000 Hz, voice agents maintain human-parity naturalness even under 10% packet loss, avoiding the robotic distortion caused by unbuffered frame dropping.

Cartesia Audio Pipeline Guardrails

• Barge-In Flush Signal: When a user begins speaking during audio playback, the application must transmit a context cancellation packet to Cartesia's WebSocket and simultaneously clear the local speaker buffer within 80ms to prevent voice overlap.
• Character Deduplication Cache: Frequently repeated conversational prompts (e.g. "Could you please repeat that?" or "One moment while I check your account") should be synthesized once, converted to raw audio chunks, and served from local memory cache at zero API cost.