The places where voice agents embarrass themselves in production aren’t the casual greetings. They’re the order numbers, the callback digits, the email addresses the caller has to scribble down while the agent rattles them off at machine-gun speed. Gradium AI is betting its new text-to-speech model can fix exactly that, and the company put some numbers behind the claim.
Gradium reports an 81.0% human-rated pass rate on a 500-sentence hard-case set spanning five languages, beating Cartesia Sonic 3.6 at 75.1% and ElevenLabs v3 Conversational at 65.4%. Fish Audio and Inworld trail further back at 49.5% and 46.5%. The evaluation set is open-sourced on Hugging Face under CC BY 4.0, covering 10 criteria — spelling, acronyms, alphanumeric tokens, dates, regular and floating numbers, emails, plus composite scenarios like IT tickets and claims calls. Scoring is strict: a single dropped digit fails the sentence. Independent native-speaker raters, loudness-normalized audio, and a cap of 40 comparisons per rater keep the methodology honest.
Latency tells the other half of the story. On Coval’s TTS benchmark, Gradium posts 216 ms P50 time to first audio — 170 ms faster than the model it replaces — with a 30 ms p75-p25 interquartile range across 480 runs. That consistency matters more than the headline number. Cartesia sits at 454 ms median with a 165 ms spread, which is over a third of its own median. Callers don’t experience averages; they experience tail turns. Gradium isn’t the fastest overall — Inworld TTS 2 claims 166 ms — but the pitch is joint position: lowest hard-case failure rate at sub-250 ms first audio with minimal variance.
The model went live as the default across Gradium’s API and Studio on August 31, 2026. No migration required, and existing custom voice clones keep working unchanged. Teams can pull the Python SDK, point at the WebSocket endpoint, and reuse their existing voice IDs. The caveat, which Gradium doesn’t hide, is that this is a vendor-run benchmark — though open-sourcing the 500-sentence set at least lets competitors and skeptics run their own numbers. Given how much money flows through voice agents that still fumble credit card digits, a model tuned specifically for the boring, high-stakes parts of a call feels overdue.
