Open Apriel Voice 27B model takes speech directly and calls tools

ServiceNow researchers released Apriel Voice 27B Preview, an open model built on Gemma 4 26B-A4B that listens to speech directly and replies or makes tool calls without first producing a transcript or using a separate dialogue model. It still accepts text and images, and spoken replies come from a separate, swappable text-to-speech system. In the authors' tests it completed 65.51% of tasks on the τ-Voice customer service benchmark with a mean latency of 1.91 seconds, against 84.19% for Gemini 3.8 Live Extended Thinking and 69.82% for GPT-realtime-2.1, and well ahead of the other open speech-conditioned models they compared. These are the team's own benchmarks, not independent tests, and the paper says the model prioritises English, with uneven quality in other languages. Sorted take: open voice agents are getting close to the paid ones on support-style tasks, but test your own calls first.