Cartesia text to speech is Sonic, a streaming speech model built for real-time voice agents rather than batch voiceover. It ships as sonic-3.6, speaks 44 languages, clones a voice from about ten seconds of audio, and streams audio over a WebSocket while your language model is still writing the sentence.
That last property is the reason it shows up in voice agent stacks. A batch TTS engine waits for a finished sentence. A streaming engine starts speaking on the first clause, which is what makes a phone call feel like a conversation instead of a walkie-talkie exchange.
This guide covers what the model actually ships today, which of the published numbers you can put in a latency budget, which you cannot, and how to measure the parts that only show up on your own traffic.
What Cartesia text to speech is
Sonic is Cartesia's text-to-speech family. The current generally available model is sonic-3.6, released on 27 August 2026. Cartesia's Sonic product page puts it at 44 languages, instant voice cloning from roughly ten seconds of reference audio, and sub-90 ms latency. The release changelog breaks those 44 languages down into 61 locales.
The realtime path is a WebSocket. You open a connection, send text as it arrives from your language model, and receive audio back in chunks. Cartesia's quickstart returns pcm_f32le at 44,100 Hz by default, which you will usually downsample to 8 kHz for telephony or resample for WebRTC.
One behaviour matters more than the format list. The synthesis endpoint keeps a context history and appends each new text chunk to it, so prosody carries across chunk boundaries. Without that, a sentence assembled from six token batches sounds like six separate sentences glued together, with a pitch reset at every seam.
Three things follow for anyone integrating it:
- Send text in natural clause-sized chunks, not token by token, and not one whole paragraph at a time.
- Keep one context per turn, not one per request, or you lose the prosody continuity you are paying for.
- Decide your output format at the edge of the pipeline, because every resample costs a few milliseconds and some quality.
How fast is Cartesia text to speech, really?
Cartesia's own Sonic product page publishes a sub-90 ms latency figure. That number is useful, and it is also the single most misread thing about the product.
It is a model-side claim: time from the model receiving text to the model emitting the first audio byte, measured under the vendor's own conditions. It does not include your network hop to their region, TLS setup, your audio resampling, your jitter buffer, or the telephony leg. The model documentation for sonic-3.6 does not publish an end-to-end figure at all, which is honest of it, because no vendor can know your round trip.
So treat sub-90 ms as a floor, not a budget. What you ship is:
| Segment | Who controls it |
|---|---|
| Time to first audio byte from the model | Cartesia |
| Network round trip to the synthesis region | You, via region choice and connection reuse |
| Resampling and encoding to your transport | Your pipeline |
| Jitter buffer and playout | Your transport and the carrier |
The practical move is to measure time to first audio at your own edge, at p50 and p90, and hold the vendor figure separately as a component metric. Percentile matters more than mean here: a p50 that looks fine hides the tail where callers start talking over the agent. Our voice AI latency guide walks through where the rest of the budget goes.
Two cheap wins are worth naming because most integrations miss them. Reuse the WebSocket across turns rather than opening one per utterance, and start synthesis on the first clause your language model emits rather than waiting for a sentence-final token. Both remove real time from the caller's experience without touching the model.
How good is the voice quality?
On Artificial Analysis, Sonic 3.6 held the highest Quality Elo in the text-to-speech comparison at 1282 when we read the leaderboard on 4 September 2026. That is a real, independent, third-party result, and it is also a snapshot: arena leaderboards move as models ship, so the rank you read today is not the rank you will read next quarter.
There is a deeper caveat, and it is the part the product pages skip. Speech quality scores are not portable between setups. Chiang, Huang and Lee analysed 80 papers from INTERSPEECH 2022 and found that critical details of subjective evaluation, including how listeners were recruited, filtered, instructed and paid, were routinely unreported. When they ran mean opinion score tests on three well-known TTS systems under different evaluation settings, they obtained "at least three distinct rankings of TTS models".
Read that again in a procurement context. The same three systems, the same audio, three different orderings, depending only on how the listening test was set up. A vendor's MOS and an arena's Elo are both evidence, and neither transfers to your call, your codec, your accents, or your script.
What does transfer is a measurement you run yourself on your own audio. In practice, the cartesia voice you pick, the sample rate you serve, and the domain vocabulary in your prompts tend to move perceived quality more than the model version does.
Getting the pronunciation right
Every production voice agent eventually mispronounces something that matters: a drug name, a ticker, a street, an account format read as a number instead of digits. This is the most common quality complaint on deployed agents, and it is not a model defect so much as a missing dictionary.
Cartesia handles it with custom pronunciations. You supply either an IPA transcription or a sounds-like spelling, wrapped in double angle brackets with pipe separators between phonemes, and the model substitutes it wherever the term appears. Case-insensitive matching requires Sonic 3.6 or later, and the API rejects two dictionary entries that would collide under the matching rules.
Practically, build the dictionary from real transcripts rather than from imagination. Pull the fifty terms your agent says most often, listen to each one, and only add entries for the ones that are actually wrong. A bloated dictionary is its own failure mode, because a sounds-like override that was right for one locale is wrong in another.
Speed and emotion controls exist too, and they are worth using sparingly. Pushing expressiveness up tends to trade against intelligibility on a compressed telephony channel, which is exactly the channel where intelligibility is scarcest.
Where Cartesia TTS fits in a voice agent stack
cartesia tts is one component in a chain: speech to text, a language model, then synthesis, then transport. Swapping the synthesis component changes three things at once, and only one of them is the voice.
- Latency profile. A faster model can expose a slow language model that the old TTS was masking.
- Turn-taking. Different audio timing changes when the caller thinks it is their turn, which changes barge-in and interruption behaviour.
- Content rendering. Numbers, dates, currencies and abbreviations are normalised differently by different engines, so the same script reads differently.
That third one is the quiet regression. A model upgrade that improves naturalness on the arena can start reading "Rs. 1,04,000" or "Suite 3B" in a new way, and nothing in your test suite notices unless you are listening to the audio.
How much the surrounding stack matters is measurable. Per Cekura's benchmarks, a frozen study of 7 configurations across 82 scenarios with 3 retained repeats, voice naturalness across the tested platforms ranged from 3.74 out of 5 to 4.47 out of 5, and mean response time from 1.27 s to 3.08 s.
Two caveats travel with those numbers. Providers selected their own configurations, so the spread measures whole assemblies rather than isolating any one component. And the response time is Cekura's main-agent measure, not provider-native component latency. Cartesia was not one of the 7 configurations, so there is no Cartesia figure to quote from that run. What the spread does show is that two competent teams shipping the same class of product land more than half a naturalness point apart.
How to test Cartesia text to speech on your own calls
This is where Cekura fits. Cekura tests, monitors, and self-improves voice and chat agents, which means the synthesis layer gets evaluated inside a full conversation rather than as an isolated audio clip.
Three things are worth putting under automated test the day you adopt any TTS engine:
- Snapshot the audio, not just the text. Cekura captures synthesized output per scenario so a pronunciation or prosody change surfaces as a diff, rather than as a support ticket. Our write-up on snapshot testing for voice responses covers the mechanics.
- Run the same scenarios across model versions. When
sonic-3.6succeedssonic-3.5in your config, Cekura reruns your scenario set against both and reports which turns changed, so a naturalness gain that costs you a mispronounced product name is visible before release. - Watch it in production, not only in CI. Cekura monitors live calls and scores them on the same rubric as the test suite, which is the only way to catch the accents, background noise and interruption patterns your scenarios did not imagine.
Because Cekura is a partner of Cartesia, the same lifelike voices can also drive the simulated caller, not just the agent under test. That matters more than it sounds: testing a voice agent with a robotic synthetic caller mostly measures how it handles robots. Details are on the Cartesia partner page.
No vendor benchmark, ours included, tells you what your agent will do on your own traffic. If you are moving to sonic-3.6, or deciding between two voices, the answer comes from running your scenario set against each option and reading the turns that changed. That is a bounded piece of work with a concrete output: a pass rate per scenario, a list of flagged mispronunciations, and a turn-by-turn diff between the two configurations. Book a demo and bring the ten calls you are least sure about.
Frequently asked questions
What model does Cartesia text to speech use?
Sonic. The current generally available version is sonic-3.6, released 27 August 2026, covering 44 languages across 61 locales with instant voice cloning from about ten seconds of reference audio.
Is Cartesia text to speech fast enough for phone calls?
Yes, for most designs. Cartesia states a sub-90 ms model-side latency, which is a vendor figure measured under its own conditions and excludes your network, resampling and telephony legs. Budget from your own measured time to first audio at p90, not from the published number.
Is Cartesia the best sounding TTS available?
Sonic 3.6 held the top Quality Elo at 1282 on the Artificial Analysis text-to-speech leaderboard when read on 4 September 2026, which is a moving snapshot. Published quality rankings also shift with test design, so verify on your own audio before committing.
How do I stop Cartesia TTS mispronouncing product names?
Add a custom pronunciation dictionary entry using IPA or sounds-like spelling in double angle brackets with pipe-separated phonemes. Case-insensitive matching needs Sonic 3.6 or later. Build entries from real transcripts, and keep the list short.
How many languages does Cartesia voice support?
Sonic 3.6 speaks 44 languages across 61 locales. Coverage per voice varies, so confirm that the specific voice you have chosen supports your target locale before you build around it.
