ServiceNow-AI Presents Apriel Voice 27B Preview

ServiceNow-AI researchers presented Apriel Voice 27B Preview, an open speech-language model that combines speech understanding, reasoning and tool use in one shared backbone. Built on the post-trained Gemma 4 26B-A4B checkpoint, it also accepts text and image inputs and generates responses and tool calls directly, without a separate dialogue model. The team reports deploying it on a custom serving platform for interactive voice conversations.

The researchers adapted the model to speech in two training stages: the first established a speech interface, and the second jointly adapted the projector and backbone. That stage used supervised fine-tuning, text replay, and source-dependent cross-entropy and forward Kullback–Leibler divergence supervision. Training data came from automatically constructed speech-conditioned dialogues. To support online distillation, the team used local teacher serving, overlapping teacher and student computation, and fused full-vocabulary losses.

In thinking mode, Apriel Voice scored 98.90 on Big Bench Audio and 81.13 on Speech-BFCL, and reached a 65.51% task success rate on τ-Voice with mean latency of 1.91 seconds. The paper also reports 51.27% English EVA accuracy and 74.13% experience pass@1. It was published on October 29 and submitted to Hugging Face Papers on November 2.

ServiceNow-AI researchers presented Apriel Voice 27B Preview, an open speech-language model that combines speech understanding, reasoning and tool use in one shared backbone. Built on the post-trained Gemma 4 26B-A4B checkpoint, it also accepts text and image inputs and generates responses and tool calls directly, without a separate dialogue model. The team reports deploying it on a custom serving platform for interactive voice conversations.

The researchers adapted the model to speech in two training stages: the first established a speech interface, and the second jointly adapted the projector and backbone. That stage used supervised fine-tuning, text replay, and source-dependent cross-entropy and forward Kullback–Leibler divergence supervision. Training data came from automatically constructed speech-conditioned dialogues. To support online distillation, the team used local teacher serving, overlapping teacher and student computation, and fused full-vocabulary losses.

In thinking mode, Apriel Voice scored 98.90 on Big Bench Audio and 81.13 on Speech-BFCL, and reached a 65.51% task success rate on τ-Voice with mean latency of 1.91 seconds. The paper also reports 51.27% English EVA accuracy and 74.13% experience pass@1. It was published on October 29 and submitted to Hugging Face Papers on November 2.