Cartesia vs Deepgram
Listeners rate Cartesia's Sonic 3.6 above Deepgram Flux TTS.
Cartesia also holds a tighter latency spread than Aura-2 and Flux TTS, so fewer turns stall the call.
Voice quality
Sonic 3.6 sounds more natural than Flux TTS
- Artificial Analysis plays listeners two clips of the same line without naming either model, and rates every model on the board the same way.
- Sonic 3.6 leads Deepgram Flux TTS by more than 200 points.
Provider Voice Arena Elo
Higher is better
Latency consistency
Sonic 3.5's slowest calls stay close to its fastest
- An average hides the calls that break a conversation. The spread is what a caller actually hears.
- Sonic 3.5's slowest measured call stays under half a second. Both Deepgram models run past that.
Latency Variation: Distribution of TTFA values
Lower is better
Read the latency measurements
Time to first audio in milliseconds. The box spans the middle 50% of requests; the line marks the median. Whiskers are not absolute minimums or maximums.
- Sonic 3.5: median 267ms; middle 50% 234–297ms; whiskers 140–391ms.
- Aura-2: median 304ms; middle 50% 213–424ms; whiskers 93–740ms.
- Flux: median 245ms; middle 50% 201–500ms; whiskers 100–658ms.
Tail latency
Cartesia stays ahead where it counts most
- P95 is the turn that one caller in twenty gets, so it sets how often a conversation stalls.
- Sonic 3.5 answers faster than both Deepgram models at that percentile, and it is the only one of the three whose spread stays narrow enough to be predictable.
Latency at P95
Lower is better
Accuracy
Sonic 3.6 makes fewer word errors than both Deepgram models
- Coval transcribes every clip each model generates and compares it with the script, so a lower rate means more of the script is spoken as written.
- Sonic 3.6 misreads fewer words than Flux and Aura-2, so names, numbers, and codes come through as written.
Word error rate
Lower is better
Reasons why companies choose Cartesia over Deepgram
Speech to text
Every reply starts from what the agent heard.
Cartesia3.4% streaming word error rateInk-2Deepgram6.6% streaming word error rateNova-37.4% streaming word error rateFluxOctober 2026
Consistency
At scale, the spread in latency matters more than the average.
Cartesia63ms interquartile rangeDeepgram211ms interquartile rangeAura-2300ms interquartile rangeFluxAugust 2026
Accuracy
Whether names, numbers, and terms come through as written.
Cartesia1.99% word error rateSonic 3.6Deepgram2.20% word error rateFlux2.78% word error rateAura-2September 2026
Languages
Comprehension and trust across a global customer base.
Cartesia44 languages from one model, with region-calibrated accentsDeepgram7 languagesVoice range
How far the catalogue stretches before you need your own voice.
CartesiaFull voice library, instant clone from 10 seconds, professional clone from 30 minutesDeepgram88 preset voicesDeployment
Decides whether regulated teams can use it at all.
CartesiaVPC, on-prem, on-device, or air-gapped, backed by a 99.9% uptime SLADeepgramSelf-hosting requires an Enterprise plan
Frequently asked questions
Is there a Deepgram alternative for speech to text?
Yes. Ink-2 is Cartesia's streaming speech-to-text model for voice agents. On the Artificial Analysis streaming word error rate index, Ink-2 scores 3.4%, against 6.6% for Deepgram Nova-3 Realtime and 7.4% for Deepgram Flux. Ink-2 has turn detection built in, so the agent knows when the caller has finished speaking without a separate model.
Is Cartesia's voice quality independently rated?
Yes. Sonic 3.6 ranks first of the 94 models on the Artificial Analysis Provider Voice Arena, a blind listening test where listeners compare two clips without being told which model produced either. It also ranks first on their Controlled Voice Arena, which gives every model the same set of cloned voices.
How is the latency on this page measured?
Coval is an independent evaluation platform that calls each provider's public API and publishes the full distribution of what comes back, rather than a single headline figure. The numbers here come from 478 Cartesia runs and 951 Deepgram runs across Aura-2 and Flux, measured over the same period.
Why does the spread matter more than the average?
A caller never experiences an average. They experience the turn they are on. A model that is usually quick but occasionally slow produces calls where the caller starts talking over the agent, and those are the calls that get abandoned. Holding the slow end close to the fast end is what makes a voice agent feel like a conversation.
Can I run Cartesia on my own infrastructure?
Yes. Cartesia deploys on-prem and air-gapped, which teams in healthcare, finance, and government use to keep audio inside their own network, backed by a 99.9% uptime SLA.
How long does it take to switch from Deepgram?
One to two weeks for most teams. Cartesia is integrated into the major voice agent platforms and every model is on the public API, so there is no platform to move onto and nothing to unpick later.
Get started today
Talk to an expert.
Connect with a member of our team and learn how Cartesia can help you build world-class voice experiences.
Start building.
Access our models via API and bring a voice agent into production in minutes.