Apriel Voice 27B Preview is an open speech model from ServiceNow-AI that handles listening, reasoning and tool calls in one shared backbone.
Architecture: Built on the post-trained Gemma 4 26B-A4B checkpoint, it processes speech directly and produces both replies and tool calls, with no separate dialogue model. It still takes text and image input.
Training: Two stages. The first sets up the speech interface, and the second jointly adapts the projector and backbone using supervised fine-tuning, text replay and online distillation.
Results (thinking mode): 98.90 on Big Bench Audio, 81.13 on Speech-BFCL, 65.51% task success on τ-Voice, 51.27% English EVA accuracy, 74.13% experience pass@1, and a mean latency of 1.91 seconds.
Deployment: The authors run it on a custom serving platform for interactive voice use.
Sign in to comment.
TL;DR of the paper, generated automatically.
Apriel Voice 27B Preview is an open speech model from ServiceNow-AI that handles listening, reasoning and tool calls in one shared backbone.
One model, speech in, tool calls out.