Gradium changed its default voice model.
Gradium AI switched a new text-to-speech model across its API and Studio on August 31, 2026. The company reports an 81.0% human-rated pass rate on a 500-sentence hard-case evaluation set, giving the model a stronger showing on difficult speech inputs than several named competitors.
That score puts Gradium TTS ahead of Cartesia Sonic 3.6 at 75.1% and ElevenLabs v3 Conversational at 65.4%. Fish Audio S2.1 Pro reached 49.5%, while Inworld TTS 1.5 Max reached 46.5%—a leaderboard where handling awkward inputs matters more than producing another polished demo.
A hard-case test built around speech’s annoying details
The evaluation set contains 100 items across 10 criteria and five languages: EN, DE, FR, ES, and PT. Seven atomic criteria test spelling, acronyms, alphanumeric tokens, dates, regular numbers, large and floating numbers, and email addresses.
Three composite criteria—Orders, IT Ticket, and Claims—combine several of those challenges into one realistic agent turn. That design focuses the test on situations where a voice system must interpret structured information without turning a date, account number, or email address into audible soup.
Gradium says all generated audio came from August 2026 runs using default settings. The audio was loudness-normalized, the order was randomized, and raters were capped at 40 comparisons with an enforced break. The evaluation set is open on Hugging Face under CC BY 4.0, giving others access to the material behind the score instead of asking everyone to admire a carefully selected showcase.
The pooled results combine the 10 criteria and average performance across the five languages. That matters because a single overall percentage can hide uneven behavior, while multilingual and mixed-format prompts create more opportunities for a voice model to stumble.
Latency improves, but Gradium is not the speed champion
On Coval’s TTS benchmark, Gradium reports a 216 ms P50 time to first audio, with a 30 ms interquartile spread across 480 runs. The result is 170 ms faster than the model it replaces, which makes the default switch more than a new label attached to the same waiting time.
Gradium still does not lead the full latency chart. Inworld TTS 2 posts a 166 ms median, beating Gradium by 50 ms, while Fish Audio S2.1 Pro records 291 ms and ElevenLabs v3 Conversational reaches 329 ms.
Cartesia Sonic 3.6 sits at a 454 ms median on the same comparison. Gradium therefore lands between the fastest result and the slower named systems, pairing a 216 ms P50 with the strongest hard-case pass rate listed in the evaluation.
That combination targets voice applications where users notice both mistakes and pauses. A model that reads structured inputs correctly but waits too long feels broken; a fast model that mangles the order details has simply failed with better timing.
The rollout requires no migration
Existing users do not need to take action. Existing voices, including custom clones, keep working unchanged, and new teams can install the Python SDK, point it at the WebSocket TTS endpoint, and reuse existing voice IDs.
Gradium is also offering 1 million credits for complete hard-case failure reports on its Discord. The offer gives developers a reason to test the model outside the published set, though the report must be complete rather than a one-line complaint that the robot said something weird.
The announcement sits in a broader machine-learning conversation involving a 150k+ ML SubReddit membership base and Michal Sutter, identified as an author and data science professional. The practical takeaway remains simpler: Gradium’s new default model raises its tested pass rate, cuts first-audio latency, and keeps the integration surface familiar.
For teams already using Gradium, the change arrives without a voice migration or API detour. For everyone else, the open test set offers a way to examine the claim on the same hard cases—because “sounds good” has never been a serious benchmark.
