ServiceNow-AI Releases Apriel Voice 27B Preview Open Speech-Language Model
ServiceNow-AI introduces Apriel Voice 27B Preview, an open speech-language model built on Gemma 4 26B-A4B that unifies speech understanding, reasoning, and tool use with reported benchmarks.
Model Overview
ServiceNow-AI has published Apriel Voice 27B Preview, an open speech-language model designed to handle speech recognition, reasoning over context, sustained conversation, and tool use within a single architecture. The model is built on the post-trained Gemma 4 26B-A4B checkpoint and retains text and image input modalities alongside speech. According to the abstract, Apriel Voice directly processes speech and generates both responses and tool calls without requiring a separate dialogue model.
Training Procedure
The authors describe a two-stage training procedure to adapt the base model to speech. The first stage establishes the speech interface. The second stage jointly adapts the projector and backbone through 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. For efficient online distillation, the authors employ local teacher serving, overlapping teacher and student computation, and fused full-vocabulary losses.
Reported Benchmarks
The abstract provides the authors’ own figures for thinking mode performance. Apriel Voice achieves a score of 98.90 on Big Bench Audio, 81.13 on Speech-BFCL, and a task success rate of 65.51% on τ-Voice, with a mean latency of 1.91 seconds. It also achieves 51.27% English EVA accuracy and 74.13% experience pass@1. The model is deployed on a custom serving platform to support interactive voice interactions.
Open Questions
The abstract does not detail the model architecture beyond the shared reasoning backbone, the scale of training compute or data, ablation studies for the two-stage procedure, licensing terms for the open release, or technical specifics of the custom serving platform. No code repository is listed in the source metadata.
Source: Apriel Voice 27B Preview: An Open Model for Speech, Reasoning, and Tool Use (arXiv 2610.40738)