← Master Index
Vol. 22 Module 22.4 Lecture

Cartesia

Voice & Music AI

How This Lesson Fits the Module & Volume

ElevenLabs is the quality-and-cloning default. Cartesia is the latency-first generative TTS vendor students will see in real-time agent RFPs: Sonic-class streaming speech, state-space audio models, and conversational duplex. Capability context is still Vol. 16 TTS, real-time AI, and real-time transcription—Cartesia does not replace Deepgram; it races ElevenLabs on time-to-first-byte audio.

Next: PlayAI. Wrap any winner behind Vol. 18 streaming / WebSockets so the product can swap Sonic vs ElevenLabs without rewriting the agent.

Learning Objectives

By the end of this lesson, students should be able to:

  • Describe Cartesia as a low-latency generative TTS / voice platform (Sonic-class), not a music model.
  • Contrast Cartesia vs ElevenLabs vs PlayAI vs hyperscaler TTS on TTFB, catalog, cloning, and compliance.
  • Read a qualitative Pricing / Strengths / Weaknesses table and refuse fake token/ms benchmarks.
  • Sketch a streaming TTS call and where FastAPI/WebSocket barge-in sits (Vol. 18).
  • Keep STT as a separate bake-off even if Cartesia ships transcription features.
  • Know when hyperscaler TTS is still the right buy (DPA, regions, STT+TTS one bill).
Definition

Cartesia is a cloud voice company known for Sonic-family text-to-speech: streaming neural speech optimized for conversational latency. Public technical narrative emphasizes state-space (SSM-style) audio models rather than “just another transformer TTS.” Treat model names, Ink/STT add-ons, and SDK versions as SKUs that change—confirm live docs. Cartesia is not Suno; it does not compose songs.

Why Latency Vendors Exist

Vol. 21 voice assistants fail when time-to-first-audio exceeds a phone-call budget. Beautiful ElevenLabs audio that starts late still feels broken. Cartesia’s pitch is: stream Sonic audio fast enough that barge-in and duplex feel human. You still measure your TTFB on telephony codecs—do not paste a keynote tokens-per-second chart into an SLO.

Pricing / Strengths / Weaknesses (Qualitative)

DimensionCartesiaElevenLabsPlayAIHyperscaler TTS
Pricing postureUsage- or plan-based Sonic TTS; confirm current character/second/credit meterCharacter/credit TTS + seats; STT minutes if usedCharacter/plan TTSMetered characters + enterprise commits
StrengthsStreaming TTFB reputation; conversational/duplex fit; modern SDK/WebSocket storyVoice catalog, cloning studio, brand recognitionDialog engines + PlayHT API familiarityCompliance, regions, bundled STT
WeaknessesSmaller consumer brand/studio than ElevenLabs; cloud; model/SKU churn; STT not the historic flagshipTTFB may lose on live calls vs Sonic-classLess “latency specialist” narrativeOften less expressive generative voice

Where Cartesia Sits in the Loop

Live agent TTS

  • WebSocket / streaming bytes
  • Cancel on barge-in
  • Pair with streaming STT

Batch narration

  • ElevenLabs or hyperscaler may win MOS
  • Cartesia still usable offline
  • Do not optimize TTFB for MP3 jobs

Not this vendor

  • Music (Suno/Udio)
  • Video (Module 22.3)
  • Frontier LLM hosting (22.5)

Do

  • A/B TTFB + MOS on the same script
  • Keep voice IDs in config
  • Isolate behind your streaming gateway
  • Eval any Cartesia STT vs Deepgram/Whisper

Don’t

  • Quote fake “X ms worldwide” from marketing
  • Expose Cartesia keys to the browser
  • Assume cloning policy is looser than ElevenLabs
  • Skip Vol. 20 audio retention rules

Python: Streaming-Shaped TTS Client

Official cartesia SDK method names evolve. Pattern: API key from env, voice ID from config, stream bytes to the Vol. 18 WebSocket that already owns barge-in. Teaching stand-in below—re-read current docs before a spike.

# cartesia_tts.py — conceptual stream; confirm SDK names in live docs import os from cartesia import Cartesia client = Cartesia(api_key=os.environ["CARTESIA_API_KEY"]) voice_id = os.environ["CARTESIA_VOICE_ID"] # Stream PCM/MP3 chunks to your FastAPI/WebSocket (Vol. 18.2) for chunk in client.tts.sse( model_id=os.environ.get("CARTESIA_TTS_MODEL", "sonic"), # SKU changes transcript="I found three matching flights. Want the earliest?", voice={"id": voice_id}, stream=True, ): if chunk.audio: yield chunk.audio # push to client; honor barge-in cancel

Related Lectures

LectureRole
ElevenLabs / PlayAITTS RFP peers
TTS / real-time transcription / DeepgramSpeech stack + STT split
WebSockets / streamingDuplex transport
Voice assistants / latencyProduct + SLO
Common Misconception

“Cartesia replaces ElevenLabs everywhere.” Batch narration and cloning studio may still favor ElevenLabs. Second: Sonic latency numbers from a blog are your SLO. Third: SSM architecture means you skip MOS tests. Fourth: Cartesia is xAI Grok (different company; Groq is Module 22.5 hardware inference). Fifth: streaming TTS without barge-in cancel is “real-time.”

Knowledge Check

  1. Short Answer: What is Cartesia’s primary product posture? Answer: Low-latency generative TTS (Sonic-class streaming speech).
  2. True/False: Cartesia is a music generator like Suno. Answer: False.
  3. Multiple Choice: Pick Cartesia over ElevenLabs mainly when: (a) live TTFB/barge-in dominates, (b) you need the largest cloning studio, (c) you need on-prem Whisper. Answer: (a).
  4. Short Answer: Name the usual STT partner class even if you buy Cartesia TTS. Answer: Deepgram, Whisper, or hyperscaler ASR (any valid).
  5. True/False: Marketing millisecond claims can be copied into production SLOs without measurement. Answer: False.
  6. Multiple Choice: Barge-in cancel lives in: (a) your WebSocket/gateway, (b) the CSS theme, (c) Suno credits. Answer: (a).
  7. Short Answer: Which Vol. 16 lecture covers streaming transcription? Answer: Real-time transcription (16.2) or real-time AI (16.1).
  8. True/False: Groq (LPU inference) is the same company as Cartesia. Answer: False.
  9. Multiple Choice: Qualitative weakness vs ElevenLabs often includes: (a) smaller studio/brand catalog, (b) infinite on-prem, (c) free unlimited cloning legally. Answer: (a).
  10. Short Answer: Which Vol. 18 topic carries streamed audio to the browser/phone? Answer: Streaming or WebSockets.

Key Takeaways

  • Cartesia = Sonic-class low-latency TTS for live agents, not a music or LLM host.
  • A/B TTFB and MOS against ElevenLabs/PlayAI/hyperscalers on your audio path.
  • Keep STT independent; wrap TTS behind Vol. 18 streaming.
  • Never invent latency or price numbers from this page.
  • Next: PlayAI—PlayHT lineage conversational TTS.
Trainer’s Guide

Lab: Same confirmation sentence streamed from Cartesia (or recorded Sonic sample if no key) vs ElevenLabs. Students log TTFB, note barge-in cancel design on a Vol. 18 WebSocket sketch, and write a one-page vendor split. Grade: no fake ms tables, env-based keys, STT vendor named explicitly.

Recap: Cartesia competes on conversational TTS latency. Measure it; do not mythologize SSM. Next generative-voice peer: PlayAI.