ServiceNow Open-Sources Apriel Voice 27B With Tool Calling

ServiceNow has released Apriel Voice 27B Preview, an open speech-language model whose paper says it "processes speech directly and generates both responses and tool calls without requiring a separate dialogue model." The paper is on arXiv and the weights are up on Hugging Face under ServiceNow-AI.

The model is "Built on the post-trained Gemma 4 26B-A4B checkpoint," the authors write, and preserves text and image inputs alongside speech. They describe a two-stage training procedure: Stage 1 establishes the speech interface, and Stage 2 "jointly adapts the projector and backbone using supervised fine-tuning, text replay, and source-dependent cross-entropy and forward Kullback–Leibler divergence supervision." Automatically constructed speech-conditioned dialogues provide conversational training data, and the authors use local teacher serving plus overlapping teacher and student computation for "effective online distillation."

In thinking mode, Apriel Voice "scores 98.90 on Big Bench Audio, 81.13 on Speech-BFCL, and 65.51% task success on τ-Voice, with a mean latency of 1.91 seconds." English EVA accuracy reaches 51.27% and experience pass@1 reaches 74.13%.

The paper labels this a "Preview" release. It does not identify training-data sources, deployment hardware requirements, or coverage of languages beyond English. The authors close by saying they serve the model "through a custom platform to enable interactive voice interactions."