Cartesia
Sonic: low-latency, expressive text-to-speech for real-time voice apps
About Cartesia
Cartesia builds real-time voice models on a state-space architecture. Its Sonic text-to-speech model (Sonic-3.6 as of August 2026) streams first audio in under 90 ms per the vendor, speaks 44 languages from one model, and supports instant voice cloning from about 10 seconds of audio. The same account covers Ink speech-to-text and managed voice agents, billed with monthly credits where one character of text costs one credit.
Pros & Cons
Pros
- Among the lowest time-to-first-audio figures published by any hosted TTS vendor, which keeps conversational turn latency down
- Streaming contexts let you feed LLM tokens in as they arrive without audible seams between chunks
- Low entry price: $5/month unlocks a commercial license and instant cloning
- One model covers 44 languages, so a cloned voice can be localized instead of re-recorded
- Inline non-verbal cues and emotion controls without SSML
Cons
- Concurrency is tight on lower tiers (2 to 5 concurrent TTS requests below Scale), which limits multi-user apps
- Marketing and tooling focus on business voice agents rather than games; there is no Unity or Unreal plugin
- Credits are monthly allowances, so heavy usage pushes you to the $299 Scale plan or an enterprise deal
Key Features
Sonic text-to-speech
Streaming TTS model (Sonic-3.6) with vendor-stated first audio in under 90 ms; Cartesia says it ranks first in the Artificial Analysis Speech Arena.
44 languages from one model
A single model speaks 44 languages, and a cloned voice can be localized into other languages while keeping speaker identity.
Instant voice cloning
Clone a voice from about 10 seconds of audio; professional voice cloning is available on Startup and higher plans.
Emotion, volume, and speed controls
Generation controls for volume, speed, and emotion, plus non-verbal cues such as [laughter] inserted directly in the transcript.
Contexts and continuations
Stream text into a persistent context over WebSocket so prosody stays continuous across chunks of an LLM response.
Custom pronunciation dictionaries
Define pronunciations for proper nouns and invented terms, including IPA-style phoneme strings.
Ink speech-to-text
Streaming transcription model on the same account and credit balance.
Managed voice agents
Hosted voice agents with phone numbers and evaluations, billed per minute from prepaid agent credit.
Pricing
- 20K credits per month
- $1 prepaid agent usage
- 2 concurrent TTS requests
- Non-commercial use
- 100K credits per month
- Commercial use license
- Instant voice cloning
- 3 concurrent TTS requests
- 1.25M credits per month
- Professional voice cloning (2 voices)
- Organizations
- 5 concurrent TTS requests
- 8M credits per month
- Priority support
- 15 concurrent TTS requests
- Professional voice cloning (4 voices)
- Custom credits and volume pricing
- Custom concurrency limits
- DPAs and BAAs
- SSO and shared Slack channel
Best For
Conversational voice agents
Low-latency speech output for LLM-driven agents in support, sales, and scheduling.
Interactive characters
Real-time dialogue for AI companions and game characters where response delay breaks immersion.
Localization
Carry one brand or character voice across 44 languages.
Similar Tools
ElevenLabs
AI voice generator and voice agents platform
Murf AI
AI voice generator with 200+ realistic text-to-speech voices
TTSOpenAI
Advanced AI voice engine for natural text-to-speech
LOVO AI
AI voice generator and video editor with 500+ voices in 100+ languages
Groq
Ultra-fast AI inference powered by custom LPU silicon
Fliki
Turn text into videos with AI voices in minutes