Voice agents often break exactly where it counts most – when they need to read order numbers, callback digits, or email addresses that a caller must write down. A single missed digit or symbol forces the caller to repeat information, erodes trust, and inflates handling time. Gradium AI has addressed this pain point by releasing a new text‑to‑speech model that is now the default across its API and Studio, effective August 31 2026, with no migration required for existing voices or custom clones.
On a rigorous 500‑sentence hard‑case benchmark covering five languages, Gradium’s model achieved an 81.0 % human‑rated pass rate, outperforming Cartesia Sonic 3.6 (75.1 %), ElevenLabs v3 Conversational (65.4%), Fish Audio S2.1 Pro (49.5%) and Inworld TTS 1.5 Max (46.5%). The evaluation is strict: a sentence passes only when every alphanumeric token, email dot, hyphen, or spoken character is heard correctly; one dropped digit fails the entire utterance.
Latency is equally important for live interactions. Gradium reports a median time to first audio of 216 ms on Coval’s TTS benchmark, 170 ms faster than the model it replaces, with a tight interquartile spread of just 30 ms across 480 runs. This low variance means callers experience consistent, prompt responses rather than unpredictable tail delays.
Because the model is already live, teams can start using it immediately through the hosted API or Studio, retaining existing voice IDs and integrating via the Python SDK or WebSocket endpoint. Supported stacks include Pipecat and LiveKit, and the model covers English, French, German, Spanish, and Portuguese. An open evaluation set is available on Hugging Face under CC‑BY‑4.0, allowing teams to verify results themselves.
In short, Gradium’s new TTS delivers the accuracy needed for critical data, the low and predictable latency required for real‑time voice agents, and zero‑effort deployment for current users.
#AI #Product #VoiceAI #TTS #CustomerSupport #NoMigration