- Published on
Gradium AI Releases New Default TTS Model: 81.0% Hard-Case Pass Rate at 216 ms Time-to-First-Audio
- Authors
- Name
- DeepHorus
Gradium AI has launched a new text-to-speech (TTS) model, establishing it as the default across its API and Studio platforms as of August 31, 2026. This release directly addresses a critical failure point in voice agents, which often struggle to accurately pronounce vital information like order numbers, callback digits, and email addresses. The company claims the model achieves an 81.0% human-rated pass rate on a challenging evaluation set, setting a new benchmark for accuracy in the industry.
The evaluation was conducted using a novel, open-sourced benchmark dataset published on Hugging Face under a CC BY 4.0 license. This 500-sentence set spans five languages—English, German, French, Spanish, and Portuguese—and includes 10 criteria such as spelling, acronyms, dates, and complex alphanumeric strings. The scoring methodology was strict and human-led, requiring a sentence to pass only if every element was pronounced correctly by an independent native-speaker rater.
On this benchmark, Gradium's new model outperformed major competitors. It achieved an 81.0% pass rate, compared to 75.1% for Cartesia Sonic 3.6 and 65.4% for ElevenLabs v3 Conversational. Other tested models, including Fish Audio S2.1 Pro and Inworld TTS 1.5 Max, scored significantly lower. This highlights the new model's superior capability in handling complex, real-world voice agent tasks without requiring text normalization.
Latency is another key improvement, with the model reporting a 216 ms P50 time-to-first-audio on the Coval TTS benchmark. This is 170 ms faster than its predecessor. More importantly, the performance is consistent, demonstrated by a tight 30 ms interquartile range across 480 runs. While not the absolute fastest on the chart, Gradium positions this as the best combination of low failure rate and rapid, low-variance response times.
The deployment is designed for seamless adoption. Existing customers require no migration; the model was switched on as the default, and all existing voices, including custom clones, continue to work unchanged. New users can get started by installing the Python SDK and connecting to the WebSocket TTS endpoint using their existing voice IDs.
This release shifts the competitive landscape for voice AI. By openly sharing its rigorous evaluation set, Gradium sets a transparent standard for measuring TTS accuracy on critical data. The focus on jointly optimizing for both accuracy and low latency makes the model particularly suited for demanding customer service and transactional voice applications.
Looking ahead, the company is actively seeking feedback and offering incentives for real-world testing. Teams that report complete hard-case failures on the model's Discord channel can receive 1 million credits. This move suggests a commitment to rapid iteration based on practical deployment scenarios in the field.
