Cartesia vs Deepgram: Text-to-Speechlatency & accuracy
Measured head-to-head, not marketing: the verdict is split: Deepgram is faster on median TTFA (198 ms vs 284 ms), while Cartesia is more accurate (1.4% WER vs 1.6%). Independent data, refreshed daily, on identical inputs. Voice quality and price are not measured.
Cartesia or Deepgram: which is better?
Each provider's best model on each measured axis, last 7d. Lower is better on both:
| Axis | Cartesia (best model) | Deepgram (best model) | Measured winner |
|---|---|---|---|
| Speed: median TTFA | sonic-3.5 · 284 ms | flux-haley-en · 198 ms | Deepgram |
| Accuracy: avg WER | sonic-3.5 · 1.4% | flux-haley-en · 1.6% | Cartesia |
Not measured here: voice quality / naturalness and price. Treat those as vendor claims until measured. Full leaderboard context: the independent text-to-speech benchmark →
Every Cartesia and Deepgram model, measured
All benchmarked models from both providers on the same pinned dataset, ranked fastest-first (median TTFA), last 7d:
| Model | Provider | TTFA median | TTFA p95 | WER | Samples |
|---|---|---|---|---|---|
| flux-haley-en | Deepgram | 198 ms | 264 ms | 1.6% | 643 |
| aura-2-thalia-en | Deepgram | 268 ms | 524 ms | 1.7% | 672 |
| sonic-3.5 | Cartesia | 284 ms | 425 ms | 1.4% | 672 |
| sonic-3.6 | Cartesia | 288 ms | 372 ms | 1.4% | 672 |
Data by Coval
Coval sells voice-agent evaluation infrastructure. Numbers are a rolling 7d aggregate on a pinned dataset under production-realistic conditions. Openbenchmarks mirrors the results with attribution; methodology and runner are open-source.
Cartesia vs Deepgram: common questions
Is Cartesia faster than Deepgram for text-to-speech?
On the current 7d window, Cartesia's fastest model (sonic-3.5) has a median TTFA of 284 ms, vs 198 ms for Deepgram's fastest (flux-haley-en), so Deepgram is faster on measured median latency. Distributions matter too: p95 is 425 ms for Cartesia vs 264 ms for Deepgram.
Which is more accurate, Cartesia or Deepgram?
By Word Error Rate: Cartesia's best model (sonic-3.5) averages 1.4%, vs 1.6% for Deepgram's best (flux-haley-en). Cartesia leads on measured accuracy. Lower is better; WER is measured on identical inputs under production-realistic conditions.
Which should I pick, Cartesia or Deepgram?
The measured verdict is split: Deepgram is faster (median TTFA), Cartesia is more accurate (WER). Pick by your constraint: realtime workloads care about latency; fidelity-critical workloads care about WER. Note voice quality / naturalness and price are not measured here.
Where does this data come from?
Coval sells voice-agent evaluation infrastructure. Every model runs the same pinned dataset under production-realistic conditions. The table is a rolling 7d aggregate, not a single job; Openbenchmarks mirrors it with attribution. Methodology and runner are open-source (Apache-2.0).