Reported information
ServiceNow opens Apriel Voice 27B, which hears speech but answers in text
ServiceNow published Apriel Voice 27B Preview on Hugging Face under the MIT licence on 29 October 2026. The model understands spoken requests and produces text responses and tool calls; it does not generate spoken replies itself.
It combines a Whisper encoder with a Gemma 4 backbone and retains text and image inputs. A speaking assistant needs an external speech synthesiser and turn-handling logic.
The model card warns of uneven performance across languages, problems with noise and long recordings, and potentially incorrect tool calls. Team-reported scores do not guarantee the same performance in another application.