Gradium’s new default voice model reads an order number back correctly 81% of the time

Gradium made a new text-to-speech model its default on 31 August 2026, tuned for the part of a voice agent that actually breaks: reading back order numbers, phone digits and email addresses. On a 500-sentence hard-case set across five languages it passes 81.0%, against 75.1% for Cartesia Sonic 3.6 and 65.4% for ElevenLabs v3 Conversational, and reaches first audio in 216ms at the median — 170ms faster than its previous model. The evaluation set is published on Hugging Face under CC BY 4.0 and the comparison also appears on an independent voice benchmark.

What it means Because the eval set is open, you can re-run this against your own scripts instead of taking a vendor chart on trust.